Computational Aesthetic Measurement of Photographs Based on Multi-features with Saliency

Yimin Zhou, Yunlan Tan, Guangyao Li

Abstract: Based on the existent computational aesthetic measurements, we present a new approach that combining both saliency region detection and extraction with a feature set in line with the principle of human vision. We first extract the saliency region using frequency-based method, then extract 53 features from both local and global regions, and select top 15 features which can determine the best aesthetic value. We run both SVM classification & regression and CART as well as linear regression on the filtered dataset. The experiments show a meaningful result of an accuracy above 70%.