网页视觉注意的相关研究

发布时间：2018-12-27 18:36

【摘要】：众所周知,我们人类的眼睛每天都会接收大量的视觉信息,凭借着我们高效的视觉注意系统,我们可以对接收而来的大量视觉信号进行选择与过滤,删除其中冗余的部分,将最重要的信息通过神经系统传递给我们的大脑,以便进行下一步的处理。在现代,众多科研工作者也正在试图将人类这种高效的视觉注意机制应用到电子计算机中,目的是使得电子计算机能够模拟人类的视觉注意系统,从而帮助人类进行更高级别、更加智能化的处理任务。目前,针对自然场景的视觉注意预测模型被相继的提出来,但是针对网页这种非自然场景的视觉注意方法却鲜有人研究。由于网页经常是图片、文本、商标、广告等的结合体,它与普通的图片相比拥有更丰富的视觉信息,另外,人们浏览网页的方式与普通图片相比也会有所不同,这使得以往的针对自然场景的传统显著性预测模型失去了效力。因此,本文将视觉注意的应用研究重点放在了网页上,并提出适用于网页的视觉注意模型。本文的主要工作有如下三个方面:首先,本文提出了一个针对网页视觉注意研究的标注数据库——WSP300 (Webpage Saliency Prediction 300),共收集网页图片300张,为了探究不同用途的网页对人眼注视区域的影响,我们选取了购物类素材共有116张,新闻类素材共有105张,社交和其它类别的素材79张。该数据库是对当前网页的视觉注意研究数据库的重要补充,并为本文之后建立的网页视觉注意模型提供了实验基础与数据支持。其次,本文提出了一种基于多特征融合的网页视觉注意的预测模型。该模型根据网页与普通图片共性与差异性,首先提出了适用于网页的自底向上的显著性特征,然后利用特征映射的方法得到独立的各特征向量,之后用机器学习的方法对这些特征向量进行训练,将提出来的特征进行有效的融合,最终得到适用于网页的视觉注意预测图(显著图)。最后,本文提出了一种基于卷积神经网络的视觉注意的预测模型。在该模型中,我们考虑了网页不仅与自底向上的驱动有关,也同样考虑到网页受到自顶向下驱动的影响。因此,我们使用全卷积神经网络(Full Convolution Network,FCN)来提取网页的高级语义信息,并与自底向上的特征进行融合。在WSP300数据库的实验上也证明了该模型的有效性。综上所述,本文一是建立了一个针对网页的视觉注意注视点的标注数据库,二是提出了两种适用于网页的视觉注意预测模型,并且通过实验表明,以上两个视觉注意模型在网页的视觉注意预测上比当前主流的视觉注意模型拥有更好的效果。
[Abstract]:As we all know, our human eyes receive a lot of visual information every day. With our efficient visual attention system, we can select and filter a large number of visual signals and remove the redundant parts. Pass the most important information through the nervous system to our brain for the next step. In modern times, many researchers are also trying to apply this efficient visual attention mechanism to electronic computers, in order to enable computers to simulate human visual attention systems, thus helping people to carry out higher levels. A more intelligent task. At present, visual attention prediction models for natural scenes have been proposed one after another, but the visual attention methods of non-natural scenes such as web pages are rarely studied. Because web pages are often a combination of pictures, text, trademarks, advertisements, and so on, they have more visual information than ordinary pictures. In addition, the way people browse the web is also different from that of ordinary pictures. This makes the traditional significant prediction models for natural scenes ineffective. Therefore, this paper focuses on the application of visual attention on web pages, and proposes a visual attention model for web pages. The main work of this paper is as follows: firstly, this paper proposes a WSP300 (Webpage Saliency Prediction database for the visual attention research of web pages. In order to explore the impact of different web pages on the eye watching area, we selected 116 materials for shopping, 105 for news and 79 for social and other categories. This database is an important supplement to the visual attention research database of current web pages, and provides the experimental basis and data support for the visual attention model of web pages established later in this paper. Secondly, this paper presents a prediction model of visual attention based on multi-feature fusion. According to the common features and differences between web pages and common images, the model firstly presents the bottom-up salient features suitable for web pages, and then obtains independent feature vectors by using feature mapping method. Then these feature vectors are trained by machine learning method, and the proposed features are fused effectively. Finally, the visual attention prediction map (salient image) suitable for web pages is obtained. Finally, a prediction model of visual attention based on convolution neural network is proposed. In this model, we consider that web pages are affected not only by bottom-up drivers, but also by top-down drivers. Therefore, we use full convolution neural network (Full Convolution Network,FCN) to extract advanced semantic information from web pages and combine them with bottom-up features. The validity of the model is also proved in the experiment of WSP300 database. To sum up, the first part of this paper is to establish a database of visual attention and fixation points for web pages, the other is to put forward two kinds of visual attention prediction models for web pages, and the experiments show that, The above two visual attention models are more effective than the current mainstream visual attention models in the prediction of visual attention on web pages.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP393.092

【相似文献】