基于稀疏性和背景先验的视觉显著性研究

CASIA OpenIR > 毕业生 > 博士学位论文

	基于稀疏性和背景先验的视觉显著性研究
	罗永康
	2016-05-25
学位类型	工学博士
中文摘要	图像是人类感知世界的视觉基础，是人类获取信息、表达信息和传递信息的重要媒介。随着当前图像获取技术的发展，人们通过各种图像采集设备方便快捷地获取大量的图像。如何从海量的图像信息中，快速有效地抽取有用的信息，成为计算机视觉领域的一个重要且急需解决的问题。人类视觉系统具有注意功能，使其能够快速地锁定关键区域，只选择进入视野的其中一部分信息用于后续更深层次的信息抽取与处理，从而获得高效的视觉感知。启发于人类视觉注意机制的显著性检测模型，可以为计算机系统实现快速高效地抽取有用的信息提供解决途径。基于视觉注意的认知理论以及生理模型，计算机领域的研究者提出了很多显著性检测模型。这些模型被广泛应用于计算机视觉、图像处理和机器人等领域。虽然这些模型取得了很不错的效果，但是其检测性能还不能跟人类视觉注意系统相媲美。视觉显著性检测仍然是一个极具挑战性的问题，还有很多问题尚未解决。图像的稀疏性和背景先验可以为显著性检测提供重要的指导信息，被广泛应用于现有的显著性检测模型中。虽然此类模型取得了很不错的效果，但是仍然存在着一些亟待解决的问题。例如，在复杂场景中，存在只利用基于稀疏性的显著性度量很难将显著区域从背景抽取出来的问题；当显著物体以较大面积接触图像边框时，易导致模型失效或者性能下降等问题。本文正是针对上述问题展开模型和算法研究，主要工作与贡献如下： 1. 针对图像稀疏性与独特性在显著性度量中的互补特性，基于认知心理学证据提出了构建显著性模型的关键四因素，并依此提出了基于图像稀疏性与独特性的显著性模型。该模型在显著值计算过程同时考虑了图像基于稀疏性和独特性的显著性度量，并采用了自适应的方法融合这两种显著性度量。该模型在不同场景下，都能获得较好的显著性检测性能，并提高了注视点预测精度。 2. 针对现有的基于边框背景先验的模型，在显著物体以较大面积接触图像边框时，存在检测方法失效或者检测性能下降的问题，提出了基于图像边框对比度和正则化流形排序的显著性模型。该模型将图像区域与边框区域的对比度用作图像前景查询，在此基础上利用正则化流形排序的方法推断图像显著值，并采用基于边框连接度的前景置信度对显著值进行修正。该模型可以较好地应对显著物体以较大面积接触边框带来的问题，以及可以有效地抑制杂乱场景中的噪声干扰。 3. 针对基于边框背景先验的显著性检测模型在利用图像区域与边框区域的拓扑关系推断显著值的过程，往往强调全局显著性，忽视局部显著性，从而导致可能错误地抑制显著物体区域的显著性或者错误地夸大背景区域的显著性的问题，提出了基于背景置信度并融合扩散过程的显著性模型。该模型利用边框连接度来度量图像区域的背景置信度，并依此计算图像基于颜色和空间的对比度的显著值，然后采用紧凑性扩散过程来增强模型对图像局部显著性的描述能力，并利用优化的方法对图像显著值赋值。该模型在检测过程中较好地兼顾了图像全局显著性和局部显著性，使其获得更好的检测性能。本文提出的模型均与当前主流模型在基准数据集上进行对比实验，实验结果验证了本文提出的模型的有效性，以及表明本文提出的模型获得了优于或者相当于当前主流模型的检测效果。
英文摘要	Image is the base of people visually perceiving the world, and it is an important media for human obtaining, representing and transferring information. With the rapid development of image acquisition technology, people can use different image acquisition equipment to obtain large-scale images conveniently and quickly. How to extract the useful information from the massive images, is an important problem urgent to be solved in computer vision. Human visual system has an attention mechanism, which enables human to select a certain subset of visual information for further processing and obtain efficient visual perception. Visual saliency models inspired by human visual attention mechanism can provide the solutions for extracting useful information in high speed and high efficiency. Inspired by the cognitive theories and physiological models of visual attention, researchers in computer science propose lots of saliency detection models, which always achieve impressive performance. These models provide wide applications in many areas such as computer vision, image processing, and robotics. However, their saliency detection performance is not comparable to human visual attention system. Visual saliency detection is still a challenging question, and there are lots of problems that have not been solved. The sparsity and background prior of image can afford importance and constructive information for saliency detection, and they are widely used in many existing saliency models. Although these saliency models obtain superior performance, they still have some problems urgent to be solved. For example, in complex scenes, measuring the saliency only with sparsity cannot make the salient region pop out from the image; when the salient object broadly touches the image boundary, these models are fragile and may fail. The major contributions of this thesis are: 1. Considering the complementary property of sparsity and distinctness in saliency detection, we introduce four key factors based on psychological evidence for saliency detection, and propose a novel saliency model based on sparsity and distinctness. The proposed model considers the sparse saliency and the distinctive saliency simultaneously, and integrates these two kinds of saliency measurement with an adaptive method. The proposed model work well in different scenes and improve the accuracy of predicting human fixations. 2. To address the problem of the existing saliency models based on boundary background prior, i.e. when the salient object broadly touches the image boundary, these models are fragile and may fail, we propose a novel saliency model which bases on boundary contrast with regularized manifold ranking. The proposed model inferences the saliency by regularized manifold ranking method with the foreground queries based on boundary contrast, and refines the saliency with the foregroundness based on boundary connectivity. The proposed model can deal with the problem of salient object broadly touches the image boundary, and suppress the noise in the clutter image regions. 3. Furthermore, to address the problem of that the existing saliency models usually consider more global saliency than local saliency in the process of calculating saliency with topological relations between the image regions and the boundary regions, we propose a new saliency model based on backgroundness, combining with a diffusion process. First, the proposed model measures the image backgroundness with boundary connectivity, and uses color and spatial contrast based on backgroundness to measure the saliency. Second, the proposed model adopts a compact diffusion process for strengthen the local saliency descriptive ability of the model, and assigns the saliency value with optimization method. The proposed model can consider the global saliency and local saliency simultaneously, which make it achieve better performance. We compare the proposed models with state-of-the-art models on the benchmark datasets. The experimental results demonstrate the effectiveness of the proposed models, and show that they all achieve comparable or better performance than stat-of-the-art models.
关键词	视觉显著性模型注视点预测显著物体检测稀疏性背景先验
语种	中文
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/11602
专题	毕业生_博士学位论文
作者单位	中国科学院自动化研究所
第一作者单位	中国科学院自动化研究所
推荐引用方式 GB/T 7714	罗永康. 基于稀疏性和背景先验的视觉显著性研究[D]. 北京. 中国科学院大学,2016.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
基于稀疏性和背景先验的视觉显著性研究.p（14909KB）	学位论文		限制开放	CC BY-NC-SA