Comparative Approach for CNN Versions for Facial Emotion Recognition by Varying Layers and Datasets

Facial emotion recognition has multiple applications in human computer interface, marketing and is becoming a demanding research field. Multiple researchers have various approaches for facial expression based on emotion recognition, In this work, we propose Convolution Neural Network(CNN) for emotion recognition, because faces are visual representation of emotions and CNNs are useful for working on images. This article presents comparison among the 3 CNN models (CNN 1, CNN 2 and CNN 3). The work is designed to verify the improvement in performance of the Convolutional Neural Networks by varying the datasets and/or CNN layers for emotion detection. In this work we performed tests on two datasets and three different versions of CNN models. The datasets variation done is for FER 2013 and the RAF-DB dataset. The performance parameters are checked for variations for various parameters for comparison like Precisions, Support, Confusion Matrix, Recall and f1-score. For FER 2013 dataset, the CNN 1 model has 53% accuracy, CNN 2 model has 56% accuracy and CNN 3 model has 60% accuracy, which leads to a conclusion that each model's results vary either according to the layer design. The dataset variation experimentation shows RAF-DB dataset outperforms for all three versions of CNNs. For RAF-DB dataset, the CNN 1 model has 61.5% accuracy, CNN 2 model has 62% accuracy and CNN 3 model has 64% accuracies. Thus, varying dataset improved performance parameters. The important takeaway of the paper is the comparative approach for layer wise CNN design, dataset variation, and finally the positive effect of adding dropout in CNN 3 resulting in increase in performance factors.