Structured low-rank inverse-covariance estimation for visual sentiment distribution prediction

Visual sentiment analysis has aroused considerable attention with the increasing tendency of expressing sentiments via images. Most previous studies mainly focus on predicting the most dominant sentiment categories of images while neglecting the sentiment ambiguity problem caused by the fact that th...

Full description

Saved in:
Bibliographic Details
Published in:Signal processing 2018-11, Vol.152, p.206-216
Main Authors: Liu, Anan, Shi, Yingdi, Jing, Peiguang, Liu, Jing, Su, Yuting
Format: Article
Language:English
Subjects:
Image sentiment
Label distribution learning
Low-rank
Structured sparsity
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Visual sentiment analysis has aroused considerable attention with the increasing tendency of expressing sentiments via images. Most previous studies mainly focus on predicting the most dominant sentiment categories of images while neglecting the sentiment ambiguity problem caused by the fact that the image sentiments elicited from viewers are very subjective and different. To tackle this problem, many research efforts have been devoted to visual sentiment distribution prediction, in which an image is characterized by a distribution over a set of sentiment labels rather than a single label or multiple labels. In this paper, we propose a structured low-rank inverse-covariance estimation algorithm for visual sentiment distribution prediction. The proposed model incorporates low-rank and inverse-covariance regularization terms into a unified framework to learn more robust feature representation and more reasonable prediction model simultaneously. In particular, low-rank regularization term plays a pivotal role in capturing the low-rank structure embedded in data and seeking the lowest-rank representation of samples in a latent low-dimensional subspace. Inverse-covariance regularization term is introduced to enforce the structured sparsity of regression coefficients by taking the multi-output structure into account. We also develop an alternative heuristic optimization algorithm to optimize our objective function. Experiment results on three publicly available datasets, i.e., Emotion6, Flickr_LDL and Twitter_LDL, using six measurements demonstrate the superior prediction performance compared with state-of-the-art algorithms.
ISSN:0165-1684
1872-7557
DOI:10.1016/j.sigpro.2018.06.001