Visual search using deep learning : group emotion recognition using deep learning

Abstract

Deep learning is a massive research field due to its possible imitation of human behaviours to automate and speed up processes. With a sea of applications such as image recognition, natural language processing and autonomous vehicles, this project would focus on the group emotional field. Other than communicating through words and actions, emotions could also convey messages. This project aims to train a deep learning network to classify group emotions inferred from images as negative, neutral or positive. The objective of this project is to work towards text-based image retrieval for a personal gallery. A dataset was requested online to train two deep learning architecture models, VGG and ResNet. These trained models would be able to recognize different features from images and then classify them. Results from the models would be combined using an ensemble to have the final classification. A test dataset, which mimics a personal gallery, was created to test the performance of the ensemble network. Using ablation studies, there is further analysis of the ensemble network to identify the best model which would be selected to construct a GUI application. There would also be experimental results and discussions that would be shown in this report.