Browsing by Author "Fontes, Luis Miguel Aires Marques da Silva"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- Scaleddown EfficientNet for Facial Expression RecognitionPublication . Fontes, Luis Miguel Aires Marques da Silva; Pinheiro, António Manuel GonçalvesAutomated Facial Expression Recognition (FER) has attracted significant interest over the last few years in the field of computer vision, with a wide range of applications ranging from humanmachine interaction to security, education, and healthcare. Transfer learning techniques allow to use models trained for pattern recognition in a given domain/ task and use their learned features to enhance their performance in a different task. This has become the fastest and most adopted way to train Convolutional Neural Networks (CNNs) for the task of FER. Although this is currently the most adopted and efficient way, it is worth considering whether a model trained from scratch for FER would yield better results. This dissertation explores the development of a specialized deep learning CNN for FER tasks without relying on transfer learning techniques. The Optuna framework was used to search for an optimal configuration of hyperparameters for a scaleddown EfficientNetV2S architecture trained from scratch, notably the depth and width coefficients affecting the number of layers and filters of the CNN. A baseline experiment was initially performed on the FER2013 dataset using the original EfficientNetV2S model architecture, with the input size adjusted to match the image size of the dataset. The results obtained in this experiment revealed performance levels comparable to human accuracy and near stateoftheart performance. The subsequent scaleddown experiment considerably increased the efficiency of the CNN by reducing the number of parameters and training time while maintaining performance levels close to those of the baseline experiment. In both experiments, the obtained global accuracy was around 66% for the FER2013 test data. The scaleddown experiment was also performed with the AffectNet test data, achieving a global accuracy of 61.3%. It should be noted that these results refer to a 7class problem on highly imbalanced datasets. For individual classes, the scaleddown experiment reached an accuracy of 85% (FER2013) and 86% (AffectNet) for test samples labeled as Happy, and 81% for Surprise test samples of AffectNet.