In this paper, we present our contribution to facial expression recognition by using image data obtained from the Light Field Face Dataset (LFFD). We compared several neural network architectures which are mainly developed around a convolutional neural network of EfficientNetV2-S and combined with different kinds of recurrent neural networks (LSTM, GRU, BiLSTM and BiGRU). Besides, we exploit different sets of sub-aperture images, each vary in terms of number of images and virtual position. The results show a significant accuracy improvement in two used configurations, depending on the sets of sub-aperture images. The first when using the model of EfficientNetV2-S in two branches configuration and composed with an LSTM. The second uses single branch model with a BiLSTM.