Visual Sentiment Recognition via Popular Deep Models on the Memotion Dataset
Bhavana Verma, Priyanka Meel, Dinesh Kumar Vishwakarma · 2024
Sentiment analysis has historically concentrated on text-based data to investigate people's views towards objects. However, given how frequently visual content is shared on social media, a strong case exists to expand sentiment analysis to include image context. This research addresses the gap through a novel approach by comparing several deep learning methods—such as CNNs and transfer learning—on the Memotion dataset. Our approach is unique in that we carefully analyze performance indicators, paying particular attention to feelings associated with the offensive class and emotions in the images. Our study explores the subtleties of sentiment analysis in visual content by assessing accuracy, precision, recall, and F1-score.By evaluating accuracy, precision, recall, and F1-score, our study delves into the nuanced aspects of sentiment analysis in visual content. Notably, we contribute valuable insights by comparing ResNet, DenseNet, EfficientNet, and Swin Transformer models, revealing that DenseNet emerges as the superior choice, showcasing advancements in image sentiment analysis techniques. This comparative analysis not only highlights the effectiveness of different models but also underscores the innovation in our methodology, providing a significant contribution to the evolving landscape of sentiment analysis in the visual domain.