视线跟踪技术的图像质量主观数据库构建及其可视化

目的 视线跟踪技术是当前最新的无干扰地获得人类视觉注意力的生物测量方法。将视线跟踪方法的原理其应用于图像质量感知研究，建立了首个基于眼动的图像质量感知数据库XJTU_ETSS。方法 应用单刺激图像质量评价法，在Tobii TX300的眼动平台上对900幅原始和失真图像进行了49人的60 Hz眼动主观测试。对含眼动的失真图像进行了可视化分析，即利用注视图、热图和自动聚类感兴趣区域图分析注意力在失真图像上的分布特点。结果 针对XJTU_ETSS眼动数据特点，利用反映群体视觉注意力共性的热图可视化方法具体对数据库中一组原始图像和其对应的6种失真图像上的热图进行了比较和定性分析，直观地显了原始图像和不同种类失真图像上视觉注意力的分布特征。结论 通过实验结果可知，视线跟踪方法能十分直观地揭示人在主观图像感知过程中的特性，这对于今后建立基于人眼注意力模型的图像客观评价算法，以及更加深入地揭示人类视觉系统的感知机制都有着现实而重要的意义。

Eye tracking based image quality assessment database and visualization method

Objective The eye tracking technique is one of the most advanced state-of-the-art physiological measurements to non-intrusively get human vision attention. Eye tracking technology is introduced and its application to the research of image quality perception is discussed. The first image quality assessment (IQA) database with eye tracking data, called XJTU_ETSS, is established for IQA research. Finally, heat map visualization is utilized on the XJTU_ETSS database to show the visual attention distribution differences within and between the original and distorted image groups. Method Single stimuli subject image assessment method is chosen on Tobii TX300 eye tracker with 60 Hz sampling rate to setup the database including 900 original and distorted images and 49 subjects. To visualize the eye tracking-based visual attention, methods as gaze plot and heat map are introduced in the paper. The distribution of values around a raw gaze data is accomplished by using an approximation to the Gaussian curve-a cubic hermite spline polynomial (cspline). Result According to the data of XJTU_ETSS, the count-based heat map, which shows the generality of the distribution of the visual attention of different subjects, is utilized to quantitatively and qualitatively analyze the different distributions of visual attention on original and distorted images. First, the total number of interest areas on original images is larger and more concentrated than that on distorted images. Second, heat maps of the distorted images, such as awgn, gb and hfn have no obvious areas of interest (AOIs). Instead, the attention of the observers is spread all over the images because those kinds of distortions added useless information to destroy the original image. Third, heat maps of the distorted images, such as jp2k, jpeg and quan don't have as many AOIs as the original images. Because those kinds of distortions subtract information to compress the image, so that the detail of the original images are lost and there are no AOIs on the detail parts of such kinds of distorted images. Conclusion According to the initial analysis on XJTU_ETSS, eye tracking techniques can intuitively reveal the features of human visual attention during original and distorted image reading. Such advantages can greatly contribute to the better visual attention-based subject image quality assessment algorithm in future and more comprehensive and thorough understanding of physiological and psychological mechanisms of human vision system.