一种显著性区域提取的新方法

A Novel Saliency Region Detection Algorithm

提出一个基于贝叶斯理论和统计学习理论的显著性提取算法.该方法基于贝叶斯理论分别阐明图像中不同特征信息、自下而上显著性和全局显著性不同位置的先验信息.本文针对特征融合问题分别使用加权线性组合Logistic模型和基于加权的非线性组合方法的正则化的神经网络来学习权值并获得所有因子.2个定位数据集的受试者工作特征(ROC)曲线的实验结果表明,我们的方法得到的显著图比其他先进的显著性模型效果更好.扩展的定量评价也证明了非线性组合优于线性组合的策略.

A visual saliency detection method based on Bayes＇ theorem and statistical learning is proposed in this paper. The method clarifies different feature likelihood and different location prior for bottom-up saliency and overall saliency in static images based on Bayes＇ theorem. For feature integration problem, a weighted linear combination method using logistic model, and a weighted non-linear combination method using regularized neural network are used to learn the weight parameters to combine all factors. Experimental results demonstrate that our method＇ s bottom-up saliency maps perform better than other state-of-the-art saliency models in two fixation data sets by the metric receiver operator characteristic （ROC） curve. And the extensive quantitative evaluation also demonstrates that the non-linear combination outperforms the linear combination strategy.