一种基于显著性区域的图像分割算法

针对传统的显著区域的分割方法存在的频谱泄露、无法达到亚像素精度的问题,提出了一种新的基于显著性区域的图像分割算法。利用Lucas-Kanade图像配准算法对大量重复出现模式的最小周期进行估计,从而准确地重建出背景模型;在此基础上,提出了"广义邻域点"的方法,设计自适应尺度的中值滤波器完成前景分量的精确分割;并提出了快速中值滤波器的设计方法,显著降低了计算复杂度。实验证明了该算法能够准确、高效地分割图像,适用于大量重复背景中前景目标的提取。

Image Segmentation based on Salient Regions

To overcome problems existing in traditional segmentation methods based on salient regions (such as spectral leakage, unreached subpixel accuracy), this paper proposes a novel segmentation algorithm based on salient regions. First, we utilize the Lucas-Kanade registration method to estimate the minimal period of a large of recurring patterns, where the background model can be reconstructed accurately. Second, “generalized neighboring pixels”are proposed, based on which the median filter with adaptive size is designed for extracting the foreground. Finally, we speed up the traditional median filter to reduce the complexity significantly. Experimental results demonstrate the algorithm deals with the segmentation task precisely and efficiently, which is suitable for extracting the foreground from images with a large of recurring patterns.