Exploring Image Visual Realism

With the rapid development of information technology and communication network, peo-ple are facing a larger variety of digital images, among them one type is computer generated(CG) images. Image visual realism is an important attribute of images that affects human perception. It is also an important factor in image quality evaluation.Visual realism is defined as the degree an image appears to people to be a photo rather than computer generated. It’s a multi-disciplinary high-level attribute, that is related to cognitive vi-sion, psychology, and other cognitive sciences. Predicting image visual realism is a challenging yet important task for the visualization and CG communities. For instance, image realism could be used as a metric for CG image quality evaluation. It is useful in manipulation of the realis-m level of computer games. Image realism could also be integrated into content-based image retrieval and image forensics.Although there is some research that more or less addressed visual realism, there is still limitations. There is no systematic analysis of visual realism, nor is there any dataset specialized for it. There is no evaluation metrics either.In this paper we systematically evaluated factors underlying human perception of visual re-alism and used that information to create an automated assessment of visual realism. We applied both bottom-up and top-down approaches.We make the following contributions. First, we established two benchmark datasets of face images and general scenes, respectively. We have published our dataset on the Internet so that it can be an extendable benchmark dataset. Second, we built empirical model of human perception of image realism of face images and general scenes. We further built empirical model of human perception of images of general scenes. Third, we identified attributes potentially related to image realism. Fourth, we designed features that are motivated by these attributes, and that can be processed by computer and used in machine learning. Finally, we proposed an attributes-motivated, automated computational model that can estimate image visual realism quantitatively.Our research differs from previous work in computer vision on image-type classification. Our method is realism-centric, focusing on estimating the realism level of individual images regardless of their types. We have following creative points.· Proposed a set of subject and object criteria for realism perception:Based on a series of human psychophysical experiments and data analysis, we found that four factors are close-ly correlated with image viaual realism, which are:image naturalness, image familiarity, semantics, and easthetics. We further designed features motivated by these attributes for automatic image realism estimation.· Integrated the research from multiple disciplines for visual realism analysis:Image visual realism is a subjective concept, which has much correlations with cognitive science and human psychology. We study visual realism from cognitive science point of view. We integrates methods from psychology and statistical analysis, like SDT, variance analysis, and significant tests. We further applied machine learning in our algorithm.· Constructed a unified benchmark dataset for quantitative realism estimation:Although there are certain research conducted on image visual realism estimation, there is no sys-tematic analysis framwork. For example, in CG community, the study of image realism is mainly reference based, and the images are limited to specific scenes. In computer vi-sion society, research is more on image type classification, instead of quantified realism estimation. Based on this background, we proposed a dataset in which each image has an empirical realism score, which laid foundation for quantitative realism estimation.