Depression Classification Based on Facial Action Unit Intensity Features Using CNN-Poolingless Framework

Depression can interfere with work, education and personal relationships and increase the risk of suicide. For this reason, there is a need for the early detection of depression in order to prevent further depression by means of ongoing treatment. Action unit (AU) intensity is associated with different emotions in facial expressions. AUs coded ordinally are more effective in facial expression recognition than binary. Depression classification is a challenging task, given that more precise action-unit intensity features and high classification accuracy of CNN methods are needed. We propose a CNN-Poolingless framework for depression classification based on 14 AU intensity features in facial expressions. An optimised lightweight CNNs using Grid Search without pooling layer framework was used to classify depression. Pooling layer states can remove important AU intensity features and affect classification results. GridSearch used to obtain the best hyperparameter values include number of filters, batch size, optimizer, and learning rate. The experiments were conducted using the DAIC WOz data set which prepared in 12 data variations. For model robustness test, we also test the model in the CASME II dataset. Experimental results show that eliminating the pooling layer is an idea worth considering because in different experimental scenarios and different data, the performance of the proposed framework is better than including the pooling layer. And when compared with the performance of other models, the proposed model has better performance. The proposed framework attained an accuracy score of 0.988, a loss score of 0.031, and an F1-Score of 0.991, indicating good Precision and Recall.