Image quality tendency modeling by fusing multiple visual cues

Yiyang Yao,Tengfei Wu,Jun Li
DOI: https://doi.org/10.1007/s11042-017-4821-8
IF: 2.577
2020-01-01
Multimedia Tools and Applications
Abstract:Flickr is a photo/video hosting website having over 87 million users. Over 3.5 million new photos are uploaded daily. Currently there is no tool to organize these huge number of users in terms of their aesthetic tendency. Although Flickr allows users to manually join in different groups, they are laborious to maintain and should be updated when new users are added or removed. In this paper, we propose to classify Flickr users into a series of circles. Each circle contains users with similar aesthetic tendencies. We observed that: 1) an idea aesthetic model should be flexible, this is because different visual features are usually representative toward different data sets, and 2) the numbers of photos from different Flickr users vary remarkably. In this work, a novel probabilistic topic model is proposed to describe the aesthetic interest of each Flickr user as a distribution in the latent space. Afterward, an affinity graph is constructed to describe the similarity in terms of aesthetic interests among Flickr users. It is apparently that, densely linked Flickr users are with similar aesthetic interests. Therefore, these users are grouped into different Flickr circles based on an efficient dense graph discovery. The pipeline of the proposed Flickr circles discovery is fully automatic. Extensive experiments shown that our proposed method accurately mines Flickr circles from 60,000 Flickr users. Both BER and F 1 scores demonstrated that our Flickr circle discovery algorithm outperform 11 other popular community detection algorithms. Furthermore, the adopted regularized topic model is more effective than several well-known topic models such as pLSA and LDA.
What problem does this paper attempt to address?