基于对象性和多层线性模型的协同显著性检测

Co-saliency detection based on objectness and multi-layer linear model

针对背景环境复杂的图像组中协同显著性检测的共显性目标混乱不一致、准确率低的问题,提出了一种基于对象性和多层线性模型的图像协同显著性检测方法。首先通过显著性先验和对象性概率加权的背景引导因子BGO计算图像间显著性引导传播的显著值;然后设计了一种局部区域特征计算图像内显著值,并使用图像的hu矩的零、一阶和二阶矩对两阶段显著值进行整合;最后通过多层线性模型自适应地融合各个显著图得到最终结果。实验结果表明:本文算法分别在iCoseg和MSRC两个数据集上的平均精度达到了87.80%和83.50%,在其它实验指标上的评估结果也有明显提高,增强了算法的适应能力。

To address the confusion and low accuracy of salient objects in co-saliency detection for image groups with complex environments,we proposed a co-saliency detection model based on objectness and a multi-layer linear model.First,we calculated the inter-saliency values using the background guidance factor weighted by saliency prior and objectness probability.We then designed a local region feature to calculate the intra-saliency values.The zero,first,and second Hu moments of the image were used to integrate the two-stage saliency values.Finally,saliency subgraphs were adaptively fused using a multi-layer linear model.Experimental results reveal that the AP scores of the proposed algorithm are 87.80%on iCoseg datasets and 83.50%on the MSRC dataset.Results from the evaluation of other experimental indicators are also improved significantly.The detected salient objects are more accurate and the adaptability of the algorithm is enhanced.