球面图像的SLIC算法

Spherical Image-based SLIC Algorithm

简单线性迭代聚类(simple linear iterative clustering,SLIC)超像素分割算法可以直接用于等距柱状投影(equirectangular projection,ERP)的球面图像,但是投影所造成的球面数据局部相关性破坏,会导致SLIC算法在ERP图像的部分区域无法生成合适的超像素分类,从而影响该算法的性能.为解决这一问题,首先对ERP格式的球面图像进行重采样,生成球面上近似均匀分布的球面像元数据;然后在保持球面图像数据局部相关性的基础上,将重采样数据重组为一个新的球面图像二维表示;并基于此二维表示,将球面数据的几何关系整合到SLIC算法中,最终建立球面图像SLIC算法.针对多组ERP图像分别应用SLIC算法和本文提出的算法,对比2种算法在不同聚类数量下的超像素分割结果.实验结果表明:所提出的球面图像SLIC算法在客观质量上优于原SLIC算法,所生成的超像素分割结果不受球面区域变化影响,且轮廓闭合,在球面上表现出了较好的相似性和一致性.

Simple linear iterative clustering(SLIC)can be applied directly to spherical images in equirectanguler projection(ERP)form.However,the damage of the correlation of the spherical data caused by projections leads to inappropriate superpixels in some areas of the ERP image,which impacts on the performance of the algorithm.To address this issue,resampling for ERP images was first applied to generate spherical image elements which are nearly uniformly distributed on the sphere.Then we rearranged those resampling data to form a novel 2D representation of a spherical image while maintaining the local correlation of spherical image data.Based on such a 2D representation,we integrated the geometrical relations of the spherical data into the SLIC algorithm and finally built a spherical imagebased SLIC algorithm.The SLIC algorithm and the proposed algorithm were respectively applied to several groups of ERP images,and the superpixel segmentation results with different clustering numbers generated by the two algorithms were compared.The experiments suggest that the proposed spherical image-based SLIC algorithm outperforms the original algorithm in terms of objective quality,and can generate superpixels without the effect of the variation of the regions on the sphere.The generated superpixels also have closed-contours and present better similarity and consistency on the spherical surface.