多核学习框架下多线索融合的显著性区域检测算法

Saliency region detection method based on multi cues fusion and multi kernel learning

针对自底向上的显著性检测算法中存在的底层特征表达力弱、产生的显著图存在噪声等问题,提出了一种多核学习框架下多线索融合的显著性区域检测算法.首先,提出全局对比度和层次空间两种自底向上的显著性线索,产生的显著图为弱显著图;其次,以弱显著图为基础,得到正样本和负样本,每个样本用颜色和纹理特征表示;最后,在多核学习框架下进行多线索融合,得到自上而下的强显著图.在公开数据集上进行的实验结果表明,文中算法优于流行的显著性检测算法,可得到更高的准确率和查全率.

The bottom-up saliency detection methods suffer from two problems ： weak discriminative ability of image low-level features and noisy saliency maps. To overcome the problems, we proposed a multi kernel saliency detection framework in which multi cues are fused. First, two bottom-up saliency detection methods, global contrast and hierarchy spatial, are proposed to get weak saliency maps; then the training samples are obtained from the weak saliency maps and are expressed by low-level color and texture features; finally, the bottom-up saliency cues are fused in the multi kernel saliency detection framework and a strong and top-down saliency map is obtained. The experiments on public datasets show that our method achieves better results with higher precision and recall compared to other popular saliency detection methods.