Similar visual complexity analysis model based on subjective perception

Abstract

The visual complexity analysis is a fundamental and essential attribute applied almost everywhere in visual computation. However, the existed methods mainly focus on the assessment by preference, which is identical to the statistical rating and measurement through quantization of specific metrics. Neither of them can pay attention to the flexibility of implicit logic and the influence of subjective factors in the analysis process. Therefore, the visual complexity analysis model based on individual perception is proposed, which combines objective features with subjective opinion to achieve an evaluation task that is more consistent with the visual complexity attributes of human understanding. Instead of a statistic model of rating scores, the proposed partial relation is used to represent users' subjective labels. After tahn function based pre-processing, the pair data can be learned by optimal algorithms for maximization of data margin and item dissimilarity distance. There are three visual features, Gist, Hog, and Color histogram, to depict the visual complexity globally and locally. Through data collection in a small database, the improved SVM strategy is used to train the model considering both two aspects of visual factors (objective and subjective factor). Then the model predicts the visual complexity in a vast database, and the results are highly consistent (more than 90%) with the manual evaluation through correlation coefficients such as Person, Kendall, and Spearman. The Chinese university logos and PubFig dataset are selected as research objects because of their natural visualization and latent symbolic semantics, and superior performance of the proposed model, as compared to the state-of-art algorithms, is demonstrated experimentally.