Representing Visual Complexity of Images Using a 3D Feature Space Based on Structure, Noise, and Diversity

Abstract

A 3D feature space is proposed to represent visual complexity of images based on Structure, Noise, and Diversity (SND) features that are extracted from the images. By representing images using the proposed feature space, the human classification of visual complexity of images as being simple, medium, or complex can be implied from the structure of the space. The structure of the SND space as determined by a clustering algorithm and a fuzzy inference system are then used to assign visual complexity labels and values to the images respectively. Experiments on Corel 1000A dataset, Web-crawled, and Caltech 256 object category dataset with 1000, 9907, and 30607 images respectively using MATLAB demonstrate the capability of the 3D feature space to effectively represent the visual complexity. The proposal provides a richer understanding about the visual complexity of images which has applications in evaluations to determine the capacity and feasibility of the images to tolerate image processing tasks such as watermarking and compression.