Abstract: This paper proposes an image segmentation algorithm based on invariant image features by applying manifold affine connection theory. Firstly, the algorithm divides an image into lots of small super-pixel areas by combining the local distance mapping and the watershed algorithm, which is performed in a tensor space on the Riemannian manifold. And then it finds the contact factor of the inside and outside regions between the different tangent spaces by changing the size of the closed area determined by the level set function, and gets the final target area. The specific application to medical image shows that the proposed algorithm is promising.

Key words: manifold, affine connection, level set, watershed algorithm

摘要： 通过利用流形上的仿射联络理论, 从机器学习的角度针对图像特征不变性提出了一种图像分割算法。该算法利用黎曼流形上的张量空间, 结合局部距离映射和分水岭算法, 把图像分割成很多个超像素小区域, 再通过水平集函数求取内外区域的不同切空间之间的联络因子, 进而改变由水平集函数确定的封闭区域的大小, 获得最终的目标区域。通过医学图像的具体应用, 验证了该算法的优势。

关键词: 流形, 仿射联络, 水平集, 分水岭算法