基于多尺度边缘曲率估计的图像拼接方法

Image Mosaic Method Using Multi-scale Edge Curvature Estimation

针对传统图像拼接方法仿射不变性差和对噪声敏感等问题,提出了一种基于图像边缘特征点的检测、匹配和融合方法.首先利用Canny边缘检测器提取图像的单像素边缘特征,通过边缘点邻域内的多尺度平均余弦值估计得到曲率极大值特征点,然后构建每个特征点的8×8邻域内的梯度方向直方图描述符,采用最小欧式距离与次小欧式距离比进行粗匹配,接着通过随机抽样一致性算法精匹配得到图像间的投影变换矩阵,最后采用加权平均融合算法可以得到无拼接缝隙的宽视角图像,实验表明,与基于Harris算法的拼接技术相比,该方法平均重复率高20.83%,平均定位误差低19.36%.

Aiming at improving speed and accuracy of traditional mosaic method based on Harris algorithm, a new mosaic method is proposed using edge feature points detected with curvature estimation technique. Firstly, edge feature is extracted with Canny algorithm. Those feature points from edges are measured by the curvature maximum using multi-scale average cosine estimation. Secondly, Histogram of gradient direction descriptors, established with the 8×8 neighborhood information of each feature point, areroughly matched with ratio of the minimum Euclidean distance to the second one. Thirdly, random sample consensus algorithm is used to match further. Finally, the weighted average fusion is able to generate a wide view image without mosaic gaps. Experimental results show the proposed method achieves 20.83% higher average repeatability and19.36% less average localization error than the mosaic technique based on Harris algorithm.