DNet: a depression recognition network combining residual network and vision transformer

Area of Science:

• Computer Science
• Artificial Intelligence
• Psychiatry

Background:

• Depression is a widespread mental health disorder with diagnostic and treatment challenges.
• Facial expressions offer potential biomarkers for depression severity.
• Existing methods struggle to capture subtle, localized facial changes indicative of depression.

Purpose of the Study:

• To develop an efficient deep learning network (DNet) for accurate depression identification using facial images.
• To leverage both global and local facial features for improved depression severity assessment.
• To enhance depression diagnosis through advanced feature fusion and attention mechanisms.

Main Methods:

• Proposed DNet architecture with a Feature Extraction Module (FEM) and Vision Transformer (ViT) Block.
• FEM utilizes channel and positional attention for feature maps from global and local facial images.
• Feature fusion via FPN and ViT Block for comprehensive semantic feature learning.

Main Results:

• DNet achieved Mean Absolute Error (MAE) of 6.09 and Root Mean Square Error (RMSE) of 7.85 on the AVEC2014 dataset.
• On the CZ2023 dataset, DNet obtained MAE of 6.73 and RMSE of 8.47.
• Experimental results validate the effectiveness of DNet in depression identification.

Conclusions:

• The DNet network demonstrates high accuracy in predicting depression severity from facial images.
• The proposed method effectively fuses local and global facial features for enhanced identification.
• DNet shows promise as a tool to assist in the diagnosis and management of depression.