一种结合视觉显著性的GrabCut图像分割方法

摘要

This paper presented a new segmentation method based on GrabCut combined with saliency detection.Firstly, saliency map is calculated by multi-scale regional contrast fused with center knowledge and target compactness.Meanwhile, Harris key point detection and Bayesian method can be used to improve the accuracy of the saliency map.Secondly, coarse segmentation is implemented to locate the position of object which is passed into GrabCut as the initial trimap.And the energy function which calculated by the gaussian mixture models of foreground and background is improved with saliency map.Finally, morphological operation is used to get the final result.Experiment results show that this method can achieve automatic segmentation and performance close to or even better than some existing methods.%结合显著区域检测技术和GrabCut算法,提出了一种新的图像分割方法.首先,计算融合中心知识和目标紧致度的多尺度区域对比度,通过Harris特征信息和贝叶斯模型进一步提高显著性,得到更加精确的前景检测模型;然后,对显著图进行分割获得目标的粗略位置,将位置信息用于GrabCut算法初始化,并在GrabCut初始化建模中引入显著值权重;最后,结合形态学运算改进分割的目标.实验结果表明:该方法能够实现自动分割,并且取得了接近甚至优于一些GrabCut方法的结果.