Abstract:

In this paper, a multi-scale Bayesian based saliency detection algorithm is proposed to improve the unsatisfactory accuracy of traditional Bayesian based saliency detection methods. Firstly, multi-scale superpixels are generated by segmenting the input image with superpixel segmentation algorithm（SLIC）. The background seeds are obtained according to the boundary information of superpixels, followed by the background prior evaluation with distance computation and multi-scale fusion. Secondly, the Harris operator is used to detect the corner points of the enhanced image to obtain the convex hulls. Multi-scale superpixels are fused and result in a convex hull prior. Then, the final prior is generated by combining the background prior and convex hull prior. Meanwhile, the observation likelihood probability is computed by using the color histogram. Finally, the saliency map is evaluated with Bayesian model according to the obtained prior probability map and observation likelihood probability. The experiments on public data sets of MSRA1000 and ECSSD show that the proposed algorithm achieves improved performance on both precision and recall compared to the other popular saliency detection methods.

Key words: saliency detection, multi-scale, background seeds, prior probability, Bayesian model

摘要：

针对传统基于贝叶斯模型的显著性检测算法存在准确率不理想的问题，提出了一种基于多尺度的贝叶斯模型显著性检测算法。通过超像素分割算法（SLIC）将原图分割成不同尺度的超像素，根据超像素边界信息得到背景种子，进而通过距离计算和多尺度融合得到背景先验；对原图进行颜色增强，采用Harris算子对增强图进行检测角点求得凸包，融合不同尺度下的超像素得到凸包先验；融合背景先验和凸包先验得到最终先验；利用颜色直方图和凸包计算似然概率；将最终先验和似然概率通过贝叶斯模型计算显著图。在公开数据集MSRA1000、ECSSD上与多种传统算法进行准确率和召回率对比，该算法有更好的表现。

关键词: 显著性检测, 多尺度, 背景种子, 先验概率, 贝叶斯模型