基于Kmeans++的自适应超像素投点方法

An Adaptive Initializing Superpixel Seed Points Method Based on Kmeans++

摘要:

超像素分割作为目标分割的预处理环节,能够极大地减少后续处理的数据量,对图像分割起着至关重要的作用。在大部分超像素生成算法中,初始种子点的选取都是以规则网格或随机确定,这容易导致欠分割。为了得到良好的初始种子点分布,减少种子点选取引起的欠分割,提出了一种基于Kmeans++的自适应确定超像素种子点方法,并由此改进了简单非迭代聚类算法(Simple Non-Iterative Clustering,SNIC)。实验结果表明,在不耗费大量计算成本的前提下,改进的SNIC算法相比传统算法能够得到更高的边界召回率和更低的欠分割错误率。

Abstract:

As a pre-processing step of target segmentation, superpixel can greatly reduce the amount of subsequent data processing, and plays a vital role in image segmentation. In most superpixel algorithms, seed points are sampled on a regular grid or initialized randomly, which easily leads to under-segmentation. In order to obtain a good distribution of seed point and avoid under-segmentation, an adaptively initializing superpixel seeds method based on Kmeans++ is proposed and used to improve the algorithms of SNIC. The experimental results show that the improved SNIC algorithm can get higher boundary recall rate and lower under-segmentation error rate than that of the traditional algorithm without a lot of computational cost.