Hybridization of ML & DL for Detecting Depression

Persistent sorrow and loss of interest or enjoyment in once pleasurable activities are hallmarks of major depressive illness. It is estimated that 5% of individuals globally experience depression. We propose to create prediction models based on well-known deep architectures that have the potential to produce encouraging outcomes in the identification of facial depression. Our investigation revealed that certain current solutions are deficient in two areas: "1) a small quantity of annotated depression data for deep representation learning; and 2) a significant variance in facial expressions between individuals with identical depression scores and between individuals with varying degrees of depression. We offer a deep joint label distribution and metric learning (DJ-LDML) method and model the identification of face depression as an LDL problem to address these issues. On the other hand, DML seeks to train a deep ordinal embedding using a label-aware histogram loss that is purpose-built to retain semantic similarity across video sequences (as specified by ordinal labels) for the purpose of discriminative feature learning. Overall, we were able to obtain a result from the approach of almost 97% accuracy.