A novel Image-Data-Driven and Frequency-Based method for depression detection

Major depressive disorder (MDD) is a severe mental disorder, in recent years, in order to help psychologists in clinical diagnosis, many precision medicine methods based on artificial intelligence have been proposed. This paper proposes a depression detection method based on image data that introduces frequency attention. In response to the problem of existing methods that use auxiliary information to improve performance but significantly increase computational costs, this method first applies adaptive gamma correction and deblurring data enhancement to the original data based on three universal characteristics of image data (information sparsity, inter-domain difference, and infinite granularity). Next, in accordance with the frequency richness characteristic unique to depression images, a transformer network with frequency attention is introduced for training. Finally, a pre-trained fine-tuning training strategy is used to address the problem of overfitting due to the small size of the dataset. This method achieved an RMSE of 7.36 and an MAE of 5.97 on the AVEC2013 dataset and 7.23 RMSE and 5.85 MAE on the AVEC2014 dataset. Sufficient experiments have shown that the combination of data enhancement and frequency attention in image-based depression detection is effective and promising. Code for our work can be accessed at https://github.com/DaoFaZiRan911/work01.