【作者】 张磊；

【导师】 王养利； 饶鹏；

【作者基本信息】 西安电子科技大学 ， 工程硕士（专业学位）， 2021， 硕士

【摘要】 随着计算机视觉技术的飞速发展,图像拼接算法已经逐步应用到我们生活的方方面面,地图软件,游乐场馆,医疗科技,安防领域等都可以看到拼接算法的技术应用。但复杂场景的拼接,相机的随机运动,以及各种原因产生的伪影问题,一直是图像拼接算法研究的难点,本文针对以上问题,分别从图像拼接的配准与合成两个步骤进行改进,主要研究内容与创新点如下:(1)针对单应类的投影参数会导致透视畸变,不利于多图拼接;球面投影类映射算法,对相机的运动方式要求比较严格;以网格作为配准单元的算法,忽视了图像内场景的内容等问题,本文提出了基于超像素单元的多图配准算法,以超像素为配准单元,先根据随机抽样一致性(Random Sample Consensus,RANSAC)得到的内点求解每个超像素单元的局部映射矩阵,然后再利用RANSAC筛选单平面的点,求解每张输入图像的全局相似变换,并构造了每张图自适应的非线性变换函数,使得超像素的局部单应矩阵与相似变换之间很好的过渡组合,最后选定基准图像,让每张图像相对于基准图像作运动变换,最终得到预配准的结果。能够对相机的平移运动,旋转运动,以及更为复杂的运动进行较好的预配准。(2)针对配准可能产生的小伪影,改进了动态规划的缝合线算法,提出了基于贪婪搜索的小伪影去除算法,建立重叠区域的端点之间的路径搜索模型,并结合伽马变换,改进了距离变换步骤,得到方向矩阵Dmatrix,然后结合经典的代价函数,生成了代价矩阵Ematrix,最后给出了缝合线的贪婪搜索步骤。使用该算法与传统动态规划算法进行对比,能够有效去除小伪影,效果好于传统算法。(3)针对配准可能产生的大伪影,改进了传统的图割优化缝合线算法,提出了基于超像素的大伪影去除算法,以超像素为单位,考虑超像素的颜色差异,构建了六层颜色模型,求得色差代价函数,考虑超像素的结构差异,求得每个超像素的梯度方向直方图,构建结构代价函数,引入超像素熵的代价函数,来作为调节颜色与结构之间的系数,最终,构建新的目标能量函数,将每个超像素块视为图的节点,建立图模型,使用最大流算法进行求解,得到缝合线。与传统图割算法进行对比,能够有效去除大区域伪影,效果好于传统算法。综上,本文给出了基于超像素的无伪影多图拼接算法的处理流程,并选择多个场景的图像进行拼接实验,与其他算法进行对比,验证了我们算法在不同场景下的处理伪影的效果优于其他算法。更多还原

【Abstract】 With the rapid development of computer vision technology,image stitching algorithms have been gradually applied to all aspects of our lives.Technical applications of stitching algorithms can be seen in map software,playgrounds,medical technology,and security fields.However,the registration of complex scenes,random camera movements,and artifacts caused by various reasons have always been difficult points in the research of image stitching algorithms.This article addresses the above issues from The two steps of registration and synthesis of image stitching are improved.The main research contents and innovations are as follows:(1)The projection parameters for homography will cause perspective distortion,which is not conducive to multi-image splicing;spherical projection mapping algorithms have strict requirements on the movement of the camera;algorithms that use grids as the registration unit ignore the content of the scene in the image,The same object uses different mapping parameters and other issues.This paper proposes a multi-image registration algorithm based on super-pixel units.The super-pixel is used as the registration unit,and the local mapping matrix of each super-pixel unit is solved according to the interior points obtained by RANSAC.Then use RANSAC to filter the purer points and solve the global similarity transformation of each input image.Constructs an adaptive nonlinear transformation function for each image,which makes a good transition combination between the local homography matrix of the superpixel and the similar transformation.each image is subjected to motion transformation relative to the reference image,and finally the result of pre-registration is obtained.It can perform better pre-registration of the camera’s translational movement,rotation movement,and more complex movements.(2)Aiming at the small artifacts that may be generated by the registration,the stitching algorithm of dynamic programming is improved,and a small artifact removal algorithm based on greedy search is proposed.Establish the path search model between the endpoints,and combine the gamma transformation to improve the distance transformation step to obtain the Dmatrix matrix.And then combined with the classic cost function to generate the cost matrix Ematrix,and finally the greedy search step of the stitches is given.Using this algorithm to compare with traditional dynamic programming algorithms can effectively remove small artifacts,and the effect is better than traditional algorithms.(3)Aiming at the large artifacts that may be generated by the registration,the traditional graph cut optimization stitching algorithm is improved,and a large artifact removal algorithm based on superpixels is proposed.Taking superpixels as the unit and considering the color differences of superpixels,six layers Color model,obtain the color difference cost function,consider the structural difference of super pixels,obtain the gradient direction histogram of each super pixel,construct the structure cost function,and introduce the cost function of super pixel entropy as the adjustment between color and structure Coefficients,finally,construct a new target energy function,treat each super-pixel block as a node of the graph,establish a graph model,and use the maximum flow algorithm to solve it.Get sutures.Compared with the traditional graph cut algorithm,the effect is better than the traditional algorithm.In summary,this article presents the processing flow of the artifact-free multi-image stitching algorithm based on superpixels,and selects images from multiple scenes for stitching experiments,and compares them with other algorithms to verify the processing artifacts of our algorithm in different scenarios.The effect of shadow is better than other algorithms.更多还原