TRANSFER LEARNING STRATEGIES FOR OPTIMISING FACIAL RECOGNITION ACCURACY

Human-computer interaction relies heavily on facial emotion recognition (FER), yet creating reliable models is difficult because of inter-subject variability and a lack of personalised data. In this study, we offer a method for improving the accuracy of face recognition with convolutional neural networks (CNNs) that is based on transfer learning. Initially, a generic model is trained using the massive AffectNet dataset; next, it is fine-tuned using the smaller AMIGOS dataset, which contains subject-specific data. The solution is aimed at one-dimensional emotion prediction (valence and arousal) and the Root Mean Square Error (RMSE) of 0.09 and 0.10, respectively. In addition, testing is done for alternative methods of data sampling active and passive to use and to minimize the cost of labelling and accuracy by using greedy sampling and Monte Carlo Dropout Uncertainty Estimation. This approach offers a scalable, efficient solution for personalised FER systems with real-world applications in psychological health monitoring, adaptive interfaces as well as affect-aware computing. The results demonstrate that only 20-30% of labelled personal data is adequate for near-optimal performance.