针对多幅眼底图像拼接的问题,提出一种优化的多幅眼底图像拼接方法。该方法在基于特征的层次鲁棒估计方法完成两两图像初始配准的基础上,提出一种基于重叠区域的配准修正,得到两幅图像之间的良好配准;然后建立图像之间配准关系的无向图,并将配准图像中匹配特征的几何误差和相似度作为图中边的权值,通过计算关系图的最短路径,确定基准图像;再构造基于基准图像的直接约束和基于非基准图像的间接约束的多幅图像配准整体模型,用来生成多幅拼接图像。最后在图像融合部分,提出基于距离变换的多频带融合方法,消除拼接图像中的接缝,达到平滑融合效果。实验结果表明,该方法可有效消除多幅配准的累积误差,实现多幅眼底图像输入顺序无关性的精确配准和无缝拼接。

We present an opimized method for multi-retinal image mosaicking.First, the feature-based hierachical, robust estimation method completes a pairwise registration. Then a method is proposed to achieve good registration, which is based on overlaping regions for registration correction. The graph is established by the elationship between the pairwise registrations. The geometric error and similarity are selected as the weight to find the shortest path and to decide the anchor image. The joint estimation is constructed by direct constraints from the anchor image and indirect constraints from non-anchor images. For image fusion, multi-band blending is proposed, which is based on distance transform. The stitching seam is removed and a brightness blending image is achieved. Experimental results show that the method can effectively eliminate the error accumulation, achieving a precise registration and a seamless mosaic for multi-retinal image independent of the input order.