Year of Graduation
Deep Learning for Group-Level Emotion Recognition From Visual Content
Alexandr Georgievich Rassadin
Applied Mathematics and Information Science
In this paper, we explore different approaches in group-level emotions recognition problem (GER), which refers to detecting the overall tone of an image. The dataset of images under consideration is divided into three categories, labeled as positive, neutral and negative. Our goal is to examine different deep neural network models and their combinations on the possibility of classification quality improvement. In particular, we study image features, extracted from faces and background, and their composition. By combining features extracted by models trained for scene recognition, object detection, and violence level detection, we archived an overall accuracy of 81.12 percent.