Sentiment Prediction in Scene Images via Convolutional Neural Networks

Sentiment prediction from visual content is a challenge due to the difficulty of inferring sentiment directly from the low-level visual features. Most recent researches use Adjective Noun Pairs (ANPs) as a middle level to narrow the gap between vision and sentiment. While as Convolutional Neural Networks (CNNs) is going deeper, it is becoming possible to implement rather complex mappings. In this paper, an image dataset with sentiment tags is built for training. We conduct the experiment by training 15,000 scene images on three different CNNs models, proving that deep learning can perform rather well on specific sentiment prediction task. CNNs models are also more concise and easy than those using ANPs in the aspect of data collection and engineering implementation.