Mining visual complexity of images based on an enhanced feature space representation

Abstract

An enhanced feature space to represent visual complexity of images, as would the HVS, is presented. Specifically, the ratio between the coherent and incoherent pixels in an image was used as a measure of the chromatic contributions to the visual complexity of an image. Similarly, the contrast, energy, entropy and homogeneity were modelled as the textural attributes of an image's visual complexity. Integrated into the SND feature space, these new (chromatic and textural) features facilitate a better and enhanced representation of visual complexity. Using the Corel 1000A dataset to validate the veracity of the proposal, the enhanced visual complexity space, the SND+ space, improves the capability to better represent visual complexity by a 16.7% increase in the exact correlation with a subjective (human) evaluation of the same dataset over the original SND space. Pursued further, the effective representation of visual complexity would have profound impacts in many areas of image processing and computer vision.