一种基于特征点检测和信息增益的视觉显著性提取方法

提出一种新的基于两阶段框架的显著性提取方法.在第一阶段,使用颜色增强的Harris特征点检测算法和邻域扩展规则得到图像的粗略显著性区域;在第二阶段,首先经由独立成份分析从预先决定的显著性区域提取出稀疏特征,然后根据提出的信息增益方法确定显著性,图像某点的信息增益定义为以该点为中心的圆形邻域局部熵和环域局部熵的差.最后经过与两种代表性的流行检测算法在人眼跟踪数据库上进行实验对比,验证了所提出方法的有效性以及能够在性能和计算复杂度之间折衷的灵活性.

A Method for Visual Saliency Extraction Based on Feature Point Detection and Information Gain

A novel computational saliency extraction approach is proposed under a two-stage framework. In the first stage,a pre-estimated salient region (PER) is extracted by employing color-boosting Harris feature point detection and neighborhood-extending rules. During the second stage,sparse features of the PER are firstly extracted by independent component analysis (ICA),and an information gain method is used to extract saliency,and then information gain (InfG) is defined as the difference of local entropies of two concentric regions: the neighborhood and its annular area. Experimental results demonstrate the effectiveness of the proposed method and its potential flexibility of taking a suitable compromise between performance and computational complexity.