Cognitive analysis of visual aesthetic perception

Abstract

Answering for the question what is beauty? has emerged as an issue of psychology, neuroscience and computer science during the last decade, after the long history of exploration in the field of philosophy and aesthetics. Especially, in the field of computer science, computational aesthetics pursues implementing an automated aesthetic judgement system based on the low-level features and machine learning techniques, for tangible application such as content recommendation. In this paper, as an effort of building a computational model of estimating aesthetic value of photos in content recommendation, a hypothesis that surface curvatures of objects and a place in a scene contribute to the estimation is proposed and implemented as a new visual descriptor named Local Slant Cue (LoSC) which represent catching 2.5D information which traditional local descriptors are hard to catch. Experimental results show its comparable performance just with the 30 percent of computational workload of the previous arts. However, comparative study reveals there exist a kind of glass ceiling regardless of feature selection, due to a weird attribute of the mediocre samples, which occupy an absolute majority of any given sample group, in machine learning framework. Observation to the score distributions of the mediocre group leads to the discovery of significantly high variance in consensus level among human raters for the stimuli. For quantitative validation of the observation, skewness-kurtosis map is adopted as a tool of consensus analysis and applied to a massive photo aesthetics dataset consisting of 225,000 samples, followed by the result of showing validated universality of the observation as one of four patterns, which are incompatible with Gaussianity that has been expected so far. Several computational models of visual aesthetic perception are proposed and tested from the view of how well they explain the observed patterns, finding the comparative advantage of dynamic systems model. As an effort of elaborating the idea of dynamic systems for the aesthetic perception, a new computational model named as DDM4AP (Drift-Diffusion Model for Aesthetic Perception) is proposed regarding visual aesthetic perception as a result of dynamic interaction between like factors and dislike factors. While it is concentrating to explain the wide variance in consensus level, the proposed model predicts a significantly longer latency when appreciating photos the mediocre group rather than the good or the bad, regardless of consensus level. Human subject experiments validate the prediction, supporting the model as reflecting important attributes of visual aesthetic perception in human mind. In conclusion, this study declares computational aesthetics requires new approaches of machine learning and computer vision considering dynamic interaction between two contrastive factors and selecting training data and features in accordance with such mixed data.