摘要： 本文结合物体先验和吸收马尔科夫链，提出了一种简单高效的图像显著物体检测模型。将显著性检测主要分为两个阶段，第一阶段的显著性检测中，我们结合背景先验，利用边界连通性准则，粗略筛选出位于边界的背景节点，将这个背景节点视为吸收节点，然后利用吸收马尔科夫链的随机游走方法进行显著性计算。以往的算法大都致力于基于像素或基于超像素的显著性计算，但显著物体的检测其最终目的是语义上的完整物体，因此，我们提出根据物体先验的假设，计算图像中每个像素的物体级概率，粗略估计语言上物体在图像中所处的位置，以充分利用图像信息。然后，将基于吸收马尔科夫的显著性检测结果和物体先验计算的结果相融合，得到初步的图像背景估计显著图。在第二阶段的显著性检测中，我们首先对显著图进行二值化分割，然后将分割得到的背景节点视为吸收节点，再次利用吸收马尔科夫链进行显著性计算，得到最终的检测结果。本文算法在4个公开的显著目标检测图像库（MSRA5K、MSRA10K，ECSSD和SED2）上进行了性能评测，并与现有的其他16种经典算法分别进行了定性和定量的比较。实验结果证明，本文算法产生的显著图能较好抑制背景信息且均匀地凸显目标，表现出比现有经典算法更好的PR曲线、MAE值、AUC值、F-measure值和WF-measure值。79563

关键词： 显著性物体检测；吸收马尔可夫链；背景先验；目标度量；超像素

Salient Object Detection based on Objectness

Abstract: We propose a simple but effective approach to detect salient objects by exploring both objectness prior and absorbing Markov chain。 Saliency detection is carried out in a two-stage scheme。 In the first stage, we use boundary connectivity as background prior to coarsely select the background nodes on the boundary。 And then, we take the background nodes as absorbing nodes for random walk on absorbing Markov chain。 The existing algorithms are almost based upon pixel-level and patch-level。 However, the aim of detecting salient object is an entire object。 Therefore, we propose to utilize the objectness prior to compute the probability of each pixel belonging to an object and coarsely obtain the position of the objects in images, which could utilize the image information in a better way。 And then, the objectness map and the saliency map based on absorbing Markov chain are combined to obtain the preliminary saliency map。 In the second stage, the saliency map from the first stage is segmented adaptively to binary map, and we use the refined background seeds computed by the first stage as absorbing nodes for the absorbing Markov chain to obtain the final saliency map。 Experimental results on four publicly available datasets demonstrate the robustness and efficiency of our proposed approach against 16 state-of-the-art methods in terms of PR curve, MAE, F-measure and weighted F-measure scores。 

Keywords: Salient object detection; Absorbing Markov chain; Background prior; Objectness measure; Superpixels

目录

摘要 i

Abstract i

目录 iii

1 绪论 1

1。1 研究背景和意义 1

1。2 国内外研究现状 2

1。3 本文主要内容及结构安排 5