Research Of Webpage Visual Attention

As we all know, human eyes receive a lot of visual information every day, but we can remove the redundant parts and pass the most important information to our brain with our efficient visual attention system. In modern times, many researchers also try to apply human's visual attention mechanism to computers, which aim to make computers simulate the human visual attention system to help humans processing more complex tasks.Visual attention prediction model for natural scenes has been put forward successively, but few attention has been paid to the visual attention method for Webpages. Webpages are a combination of images,text, trademarks, advertisements, etc., it has richer visual information than natural pictures, and people also behave different while browsing the webpages. Therefore, it is ineffective to predict visual attention on webpages just by employing previous attention models designed for natural images. So, our main work includes the following three aspects:First of all, our paper built a webpage saliency prediction data, which is called WSP300. In order to explore the effects of different web pages on the human fixation areas, we collected 300 web pages, among which are 116 shopping webpages, 105 news webpages and 76 social and other webpages. The database is an important supplement to our visual attention research on webpages.Secondly, this paper proposes a visual attention model based on multi-features fusion. Based on the similarities and differences between web pages and ordinary images, this paper proposes different kinds of bottom-up salient features which are applicable to web pages, and then uses feature- mapping method to obtain independent vectors, in order to achieve a flexible adaptation to various types of webpages, we employed a machine-learning framework based on our proposed features.Finally, a visual attention prediction model based on convolutional neural network is proposed in this paper. In this model, we consider the webpages not only driven by the bottom-up factors, but also the top-down factors. Therefore, we use the Full Convolution Network (FCN) to extract the high-level semantic information of web pages, and then merge it with bottom-up features. The validity of the model is proved by the experiments on our database.In this paper, we establish a webpage salieney predietion data, and two visual attention prediction models are proposed. The experimental results show that the above two visual attention models outperforms other state-of-the-art methods in webpage saliency prediction.