针对单一显著性特征无法全面表达图像显著性致使显著性检测精度不高等问题, 本文提出了一种多特征融合的显著性检测算法。算法在高层先验知识基础上, 对靠近中心的超像素设置高显著值, 利用高斯分布求解中心先验; 在底层特征上融合图像的边界稀疏、全局对比度、颜色空间分布和超级像素差异等4种显著特征, 利用类间差异最大阈值对高低层特征进行线性和非线性融合, 最终得到高质量的显著图。在MSRA-1000、SED、SOD 3个公开的数据集上进行实验, 结果表明: 本文算法融合得到的显著图边缘清晰、显著区域突出均匀, 在有效抑制背景信息的同时所得显著图像视觉感知更好, 与其他显著性算法相比查全率和查准率上至少提高3.4%。

Aiming at the problem that the single saliency feature can not fully express the image saliency,resulting in low accuracy of saliency detection,a saliency detection algorithm for multi-feature fusion was proposed. Based on the high-level prior knowledge, the background priori and the center priori were redefined.The super-pixels near the center were set to high saliency values, and the center priori was solved by using Gauss distribution. Four kinds of salient features such as image sparseness, global contrast, color space distribution and super pixel difference on the underlying features were fused, and the maximum threshold and high-low layer features of the inter-class difference were linearly and nonlinearly fused to obtain a high-quality saliency map. Experiments on three public datasets of MSRA-1000, SED and SOD show that the saliency map obtained by the fusion of the algorithm has clear edges and prominent regions, and the visual image perception is better when the background information is effectively suppressed. Compared with other significant algorithms, the recall rate and precision rate are increased by at least 3.4%.