基于多层次特征提取与匹配的视差图像拼接算法研究

发布时间：2018-01-30 18:57

本文关键词： 视差图像拼接多层次特征提取特征匹配特征分块　出处：《西南交通大学》2017年硕士论文　论文类型：学位论文

【摘要】：针对现有的视差图像拼接算法中单应性矩阵不具有全局性、计算量大,图像匹配精度不高且拼接结果有重影和结构扭曲的问题,提出了基于多层次特征提取与匹配的视差图像拼接算法。本文用基于特征分块的视差图像拼接算法提高计算效率,用多层次特征提取与匹配模型提高匹配精度,用视差图像局部优化模型消除重影和形状扭曲。主要研究内容如下:1)基于特征分块的视差图像拼接算法。首先,用图割算法将参照图像I1和目标图像I2分割成若干个具有独特性质的图像块,并将图像块编号;然后,用特征分块法确定参照图像与目标图像之间的特征匹配图像块;最后,用特征分块法计算全局单应性矩阵Hi。实验结果表明,用本文提出的特征分块算法可以快速确定图像之间的重叠区域和非重叠区域,保证了单应性矩阵的全局性,减少了计算单应性矩阵的迭代次数,提高了算法的计算效率。2)改进了多层次匹配框架,设计了一种多层次特征提取与匹配模型,对视差图像进行匹配。首先,搭建多层次特征提取和匹配框架;然后,进行多层次特征提取;最后,进行多层次特征匹配。在进行多层次特征提取时,从高分辨率层到结构层逐层进行特征汇聚,可以保证提取特征的准确性。在进行多层次特征匹配时按照"由粗到精"的策略用结构层的匹配结果指导高分辨率层的匹配,这样可以提高匹配的精度和速度。通过对比实验可以发现本文的多层次特征提取及匹配模型可以提高匹配的精度。3)设计了一种视差图像局部优化模型,消除视差图像拼接过程中产生的重影和形状扭曲。首先,采用本文提出的"结构+纹理"的方法找出最优全局单应性矩阵H,用H对目标图像I2进行变换得到预配准图像I2;然后,本文在现有的位置变化约束和相似变换约束的基础上加入块链接约束,对I2的重叠区域进行局部优化,得到局部修正后的目标图像I2;最后,用双线性插值法进行图像融合得到最终的拼接图像。对比实验证明本文算法可以较好消除重影和形状扭曲,提高了视差图像拼接的质量。
[Abstract]:In the existing parallax image stitching algorithms, the monoclinic matrix is not global, the computation is large, the image matching accuracy is not high, and the stitching result has the problems of double shadow and structure distortion. A parallax image mosaic algorithm based on multi-level feature extraction and matching is proposed. In this paper, the parallax image mosaic algorithm based on feature block is used to improve the computational efficiency, and the multi-level feature extraction and matching model is used to improve the matching accuracy. The partial optimization model of parallax image is used to eliminate double shadow and shape distortion. The main contents of this paper are as follows: 1) Parallax image mosaic algorithm based on feature partitioning. First of all. The reference image I1 and the target image I2 are divided into a number of unique image blocks by using the graph cutting algorithm, and the image blocks are numbered. Then, the feature matching image block between the reference image and the target image is determined by the feature block method. Finally, the feature block method is used to calculate the global homotropic matrix Hi.The experimental results show that the proposed feature block algorithm can quickly determine the overlapped and non-overlapping regions between images. The global property of homotropic matrix is guaranteed, the iteration times of calculating homotropic matrix are reduced, and the computational efficiency of the algorithm is improved. 2) the multi-level matching framework is improved. A multi-level feature extraction and matching model is designed to match parallax image. Firstly, a multi-level feature extraction and matching framework is built. Then, multi-level feature extraction is carried out. Finally, multi-level feature matching is carried out. In the process of multi-level feature extraction, features converge from high-resolution layer to structure layer. It can ensure the accuracy of feature extraction. According to the strategy of "from coarse to fine", the matching result of structure layer is used to guide the matching of high resolution layer in multi-level feature matching. Through comparison experiments, we can find that the multi-level feature extraction and matching model can improve the matching accuracy. 3) A parallax image local optimization model is designed. The double shadow and shape distortion generated in parallax image stitching are eliminated. Firstly, the "structure texture" method proposed in this paper is used to find out the optimal global homotropic matrix H. The target image I2 is transformed by H to get the pre-registered image I2. Then, based on the existing position change constraints and similarity transformation constraints, block link constraints are added to optimize the overlapped region of I2 and obtain the locally modified target image I2. Finally, the bilinear interpolation method is used to fuse the image to obtain the final stitched image, and the contrast experiment shows that the algorithm can eliminate the double image and shape distortion, and improve the quality of parallax image mosaic.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】