earticle

영어

In modern society, visual content like images and videos is increasingly becoming a new form of media to express users’ opinions on the Internet. As a complement to textual sentiment analysis, visual sentiment analysis intends to provide more robust information for data analytics by extracting emotion and sentiment toward topics and events from images and videos. Inspired by recent works that applied deep convolutional neural networks (CNN) to this challenging problem, we proposed a framework for image sentiment analysis with a novel deep neural network called Network in Network (NIN) which intends to improve the discriminability for local patches within receptive fields. We trained our network on a dataset consisting of nearly half a million Flickr images and minimized the effect of noisy training data by fine-tuning the network in a progressive manner. Extensive experiments conducted on manually labeled Twitter images show that the proposed architecture performs better in visual sentiment analysis than conventional CNN and other traditional algorithms.