| Popis: |
With the growing popularity of social media platforms, everyday millions of users are sharing their emotions, views, and information about notable events through visual contents especially images. Analyzing the sentiments of these user-generated images provide prevalence opportunities to various organizations for decision making. Most of the prior studies focused on hand-crafted features and traditional deep learning-based approaches. Conversely, a little effort has been employed to exploit the state-of-the-art transfer learning methods. In this paper, we propose a visual sentiment analysis system adopting the strengths of the ensemble of transfer learning models. We employ three pre-trained deep convolutional neural network (CNN)-based feature extractors including VGG16, Xception, and MobileNet to train the corresponding base classifier. Finally, a majority voting-based fusion approach is employed to determine the sentiment label. Experiments on a benchmark Twitter image dataset indicate that our method achieves top-notch improvement compared to other state-of-the-art methods. |
