Detecting a poker face

• Autor(a) principal: Sousa, Lúcia Maria Bessa de
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/40041
• topic: Face identification; Emotion analysis; Machine learning; Computer vision

Neutral facial expression recognition holds significant importance in various domains and applications. This thesis presents a comprehensive study on neutral facial expression recognition, investigating different approaches, techniques, and challenges in the field. The study employs a multi-faceted methodology. Firstly, it examines the impact of datasets, data augmentation, balancing, and creating a specialized dataset by merging existing datasets on emotion recognition. It is found that data augmentation has a significant impact on the training of the model. Subsequently, the study explores the effects of various model architectures, parameters, and training techniques to identify the most effective approach. Notably, the InceptionV3 model demonstrates the best performance with an accuracy of 72%. Additionally, the investigation includes examining different preprocessing methods and their influence on the performance of the best-performing model, InceptionV3, and a simplified CNN model. The results reveal that preprocessing techniques positively impact the performance of the simpler CNN model, leading to improvements. However, it is noteworthy that the application of preprocessing methods has a negative impact on the performance of the InceptionV3 model. These findings highlight the importance of considering the specific model architecture when choosing and applying preprocessing techniques to maximize performance and optimize results. The system implemented to validate the study, showed promising results, for the best performing model, which was subsequently tested with participants, of which 75% obtained correctly classified neutral expressions.