基于图模型的显著性检测

 2022-01-17 11:01

论文总字数:23839字

目 录

1.绪论 1

1.1 选题的目的和意义 1

1.2 国内外研究现状 1

1.3 本文的主要研究内容 1

2.基于图流形排序 4

2.1 流形排序 4

2.2 显著性测量 5

3.图形构建 5

4.两阶段显著性检测 7

4.1 背景查询排序 7

4.2 基于前景查询排序的显著性 8

4.3 算法改进 10

4.3.1 BU模型 10

4.3.2 字典构建 10

4.3.3 TD模型 11

4.3.4 加强 12

4.3.5 通过全局和局部提示来得出最终显著图 13

5.实验结果 13

5.1 MSRA-1000 14

5.2 运行 15

6.总结 15

参考文献 16

致谢 19

基于图模型的显著性检测

李欣栩

,china

Abstract:Most of the currently known bottom-up methods measure the foreground significance of a pixel or region based on its local context or contrast within the entire image,There are only a few ways to focus on segmenting the background area and thus highlighting the objects to be displayed.The contrast between salient objects and surrounding regions is not considered in this paper, but the foreground and background clues are considered in different ways.We sort image elements by comparing the similarity between image elements (pixels or regions) and foreground clues or background cues by using the method of graph manifold sorting. The significance of image elements is defined by comparing the correlation between image elements and a given seed or query.At the same time, the image is represented as a closed loop graph with super pixel as the node. In order to sort these nodes, we need to use the correlation matrix and compare the degree to which the nodes are similar to the foreground and background.In order to extract background area and foreground salient object effectively, significance detection is carried out by two-stage scheme.However, the algorithm of this paper also has some shortcomings, we also put forward some improvement measures, that is, using manifold sorting algorithm and dictionary construction method, using the foreground and background information of the image to obtain the final salient map twice before and after.Finally, the experimental results on a large benchmark database show that the proposed method has good performance in accuracy and speed.

Key words:salient object detection;super pixel;graph manifold ranking.

1绪论

当人们观察外界事物时,总会将目光放在外界事物中的一部分观察对象上,而不会放在全部观察对象上,这是由人类的视觉注意力机制造成的,人们希望计算机也可以像人眼这样只关注事物中最显著的部分并且对其他次要部分不进行考虑,因此,人们提出建立视觉显著性模型来模仿人的视觉注意力机制。

1.1选题的目的和意义

当今社会,科技进步的非常快,随着人们对计算机的研发速度不断加快和应用程度不断加深,我们能够接触到的信息也越来越多,彩色图像和视频等多媒体信息的应用范围不断的加大,它们的规模和复杂程度也不断加大,如何进行图像分析就成为一个需要我们思考和解决的问题,由于日常生活中我们接触的图像和视频都是由人类来观察的,人眼视觉系统也拥有强大的信息处理和认知能力,所以研究人眼视觉特性和找出一种途径来模仿人眼视觉特性就显得极为重要。怎样将图像和视频里含有显著物体的部分找到并提取出来已经成为人们比较关心的问题。

以往在进行图像处理时,通常会给所有的图像元素分派相同的优先级,但是通常情况下,我们只需要对图像中的某一个区域进行分析,这种相同的优先等级不仅会造成资源浪费也会使我们的分析变得复杂。从生理学和心理学角度来说,人类对已获得的图像的处理是具有选择性的,即只关注大脑认为重要的东西并对这些东西进行进一步的分析同时忽略那些图像的次要部分,这种方式有助于提高我们处理信息的速度,从计算机智能模拟的角度来说,模仿人的注意力机制来提出图片中最显著的目标可以有效地减少冗余和简化我们的分析过程。

剩余内容已隐藏,请支付后下载全文,论文总字数:23839字

您需要先支付 80元 才能查看全部内容!立即支付

该课题毕业论文、开题报告、外文翻译、程序设计、图纸设计等资料可联系客服协助查找;