Visual Sentiment Analysis for Few-Shot Image Classification Based on Metric Learning

Visual sentiment analysis is an interesting and challenging research problem. that investigates sentiment estimation from images. Most studies have focused on estimating a few specific sentiments and their intensities, using several complex convolutional neural network (CNN) models. However, sentiment estimation from a small number of images using few-shot 1-way learning has not been sufficiently investigated. This research aims to accurately estimate sentiment using few-shot 1-way learning from given images that evoke different emotions. We first introduce a visual sentiment dataset based on Plutchik's wheel of emotions, called Senti8PW. We perform a few-shot image classification using Senti8PW, where we present a highly accurate deep neural network model with a small number of parameters and convolutions. We then use Senti8PW to perform experiments using few-way 1-shot learning. We also employ the Euclidean distance and Cosine similarity as a metric of our proposed model. Each emotion is assumed to have a probability distribution. After training our deep neural network, we predict an evoked emotion for a given unknown image. We also perform experiments to compare our proposed model with existing models. The classification system of four layers of convolutions with 5-way 1-shot learning proves to be the best in terms of balancing accuracy and a number of model parameters. Thus, results demonstrate that our model outperforms existing state-of-the-art algorithms with regard to the of balance between accuracy and parameter number.