针对传统分割方法基于单个视觉线索的不足,提出一种结合两种局部边缘测度的自然场景彩色图像区域分割方法。首先,采用logistic回归模型对200幅彩色图像进行训练,建立边缘测度与对象边界的回归模型;然后,采用该模型预测图像中每个像素的边界置信度;最后,将置信度的函数以自适应权重的形式整合到均值漂移分割算法中,实现图像区域分割。近20幅图像的定量和目视对比实验结果表明,本文方法能有效地控制过分割和欠分割的产生,且具有更好的区域边界定位效果。

A regional segmentation algorithm for nature sense images is presented in this paper. The method integrates two local edge cues and aims to reply the inadequacy in single visual cue based methods. Firstly, the logistic regression model is trained by 200 color images and the regression relationship between edge cues and image object boundary is established. Then, the model is employed to predict the boundary confidence of every pixel. Finally, the confidence functions are integrated into the procedure of mean shift segmentation by calculating the adaptive variable weights. Both quantitative and visual inspections operated in about 20 images show that the algorithm has better performance in suppressing over-segmentation and under-segmentation, and tightly localizing the boundaries.