基于视频序列的图像拼接技术研究

发布时间：2018-03-20 01:25

本文选题：关键帧　切入点：图像拼接　出处：《南昌航空大学》2017年硕士论文　论文类型：学位论文

【摘要】：图像是人类获取信息的重要渠道之一。随着时代变迁,人们不再满足于普通视角图像,而是寻求宽视角图像。然而,可以直接获取宽视角图像的广角镜头相对较贵,且图像存在一定程度畸变。所以,通过普通硬件设备获取宽视角图像成为了研究热门。近年来涌现出很多基于视频序列的图像拼接算法,但是,当视频序列之间的尺度、光线、角度发生变化时,拼接效果不甚理想。为此,本文针对现有图像拼接算法中存在的问题提出改进方案,具体改进策略如下:(1)针对视频的帧间冗余问题,统计分析选择帧间隔对匹配点个数、匹配时间、误匹配率的影响,提出基于数据统计曲线的关键帧提取方法,大大减少了拼接工作量。(2)针对SIFT算法计算速度低,匹配精度不高的问题,提出了一种结合多维定标和局部纹理特征的改进SIFT匹配算法。首先,通过MDS算法对128维SIFT特征描述符进行降维,有效地提高SIFT特征匹配的计算速度。与其他降维算法相比,MDS保证了数据的几何拓扑性;然后,基于特征一致性匹配规则和比值一致性匹配规则提出一种改进的双向匹配策略;最后,分析比对原始灰度图像上匹配点对邻域的旋转不变LBP纹理特征,进一步降低了误匹配率。(3)针对融合后图像存在人为加工痕迹(拼接缝)的问题,将泊松融合引入到图像序列的拼接中。首先利用SIFT图像配准算法,获取图像序列的重叠区域,然后初始化泊松融合参数,最后通过求解泊松融合实现无缝拼接。为了解决泊松融合中颜色失真的问题,基于传统的泊松融合算法提出加权泊松融合算法。以不同场景下的实拍图像为实例进行实验仿真。仿真结果表明,与同类算法相比,该方法在匹配精度和匹配速度、融合效果上都得到一定程度的改善。
[Abstract]:Image is one of the important ways for human to obtain information. With the change of times, people are no longer satisfied with the image of common view, but seek the image of wide angle of view. However, the wide-angle lens which can directly obtain the image of wide angle of view is relatively expensive. In recent years, a lot of image mosaic algorithms based on video sequences have emerged, but when the scale between video sequences, light, When the angle changes, the stitching effect is not very good. Therefore, this paper proposes an improved scheme to solve the problems existing in the existing image stitching algorithms. The specific improvement strategy is as follows: 1) aiming at the inter-frame redundancy of video, This paper analyzes the influence of frame interval on the number of matching points, matching time and mismatch rate. A key frame extraction method based on data statistical curve is proposed, which greatly reduces the workload of stitching. An improved SIFT matching algorithm combining multidimensional scaling and local texture features is proposed. Firstly, the dimension reduction of 128-dimensional SIFT feature descriptors is achieved by MDS algorithm. Compared with other dimensionality reduction algorithms, SIFT ensures the geometric topology of the data, and then proposes an improved bidirectional matching strategy based on the matching rule of feature consistency and the matching rule of ratio consistency. Finally, we analyze the rotation invariant LBP texture features of matching points on the original gray image, and further reduce the mismatch rate. Poisson fusion is introduced into the mosaic of image sequences. Firstly, the overlapping regions of image sequences are obtained by using SIFT image registration algorithm, and then Poisson fusion parameters are initialized. In order to solve the problem of color distortion in Poisson fusion, we realize seamless stitching by solving Poisson fusion. Based on the traditional Poisson fusion algorithm, a weighted Poisson fusion algorithm is proposed. The simulation results show that the proposed method is more accurate and faster than the similar algorithms. The fusion effect is improved to some extent.
【学位授予单位】：南昌航空大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】