多高斯模型特征空间覆盖学习的海洋航摄图像分割

Aerial Ocean Image Segmentation Using Multi-Gauss Model Characteristic Space Cover Learning

本文提出了一种新的基于多高斯特征空间覆盖学习的航摄海洋图像分割方法.通过分析,发现在RGB三维色空间中,海水背景像素点的分布尽管在不同成像条件下具有不同的分布特性,但其具有的共同特性是具有集聚性,可以被一个或多个椭球所覆盖.因此,本文在色空间中基于贝叶斯最大后验概率和3δ准则对海水背景进行多高斯分布模型覆盖建模,自学习得到其高斯分布个数并建立相对应的多高斯分布模型.最后,根据上述学习结果,从航拍海洋的图像中分离出海水背景,实现航拍海洋图像中背景和目标的分割.实验证明,该方法具有良好的背景学习性能,能够准确有效地得到海水背景多高斯分布覆盖模型.基于该背景学习模型的海洋图像分割,具有较高的正确率和较低的误差,且算法花费的时间较少,具有较好的稳定性和实时性.

This paper proposes a method for aerial ocean images segmentation based on multi-gauss characteristic space cover learning.Firstly,w e analyze the distribution characters of sea background images and find that they are diversity in location,direction and geometrical morphology but clustering and can be covered by one or more spheroids.Then,w e use the multi-gauss model to describe them and get the number of gauss components adaptively based on the maximum Bayesian posteriori probability and 3δ criterion.Finally,w e segment the aerial ocean images series according to their cover learning results.The experimental results show that this method can get the cover learning model accurately and effectively and segment the aerial ocean images w ith high precision and low error in less time.