Transfer Learning for Multi-label Emotion Recognition: A Promising Approach for Biomedical Applications

Human emotions play a major role in our daily lives, influencing our thoughts, actions, and decisions. We propose a state-of-the-art, diverse, and inclusive dataset for emotional artificial intelligence (EAI) or affective computing. Our goal is to develop a system that can classify and recognize multiple emotions in images simultaneously, similar to how humans identify emotions. Using transfer learning with pre-trained deep neural networks like vgg16 added on top of it 4-layers deep neural network reaching 91.90% training accuracy with 0.28 training loss and 90.34% validation accuracy with 0.33validation loss trained on a dataset of 1000 images total being 800 images for training and 200 images for validation. The second architecture used was vgg19 with the same 4-layer deep neural network achieved 91.88% training accuracy with 0.28 training loss and 90.12% validation accuracy with 0.39 validation loss. The third and last architecture used was resnet added on top of it 4-layer deep neural network accomplished 90.15% training accuracy with 0.32 training loss and 90.90% validation accuracywith 0.33 validation loss. The results are really promising as the data has been collected reliably and accurately. Images were processed to only focus on the region of interest (ROI) in the images which is the face of the male or female that appears in the image. One of this study's primary aims is to develop a biomedicalprogram to treat children with autism, as they struggle to identify human emotions and distinguish between basic emotions such as anger, fear, surprise, happiness, sadness, and disgust.