FBANet: Transfer Learning for Depression Recognition Using a Feature-Enhanced Bi-Level Attention Network

Area of Science:

• Psychology
• Computer Science
• Artificial Intelligence

Background:

• The House-Tree-Person (HTP) sketch test is a psychological tool for assessing mental health.
• Current depression recognition from HTP sketches often relies on subjective manual analysis, limiting automation and objectivity.
• Existing automated methods using machine learning/deep learning have complex pipelines, hindering practical application.

Purpose of the Study:

• To develop a highly automated, accurate, and efficient deep learning-based method for depression recognition from HTP sketches.
• To introduce a novel one-stage deep learning architecture for improved depression detection.

Main Methods:

• A novel Feature-Enhanced Bi-Level Attention Network (FBANet) was designed, incorporating feature enhancement and bi-level attention modules.
• Transfer learning was employed, pre-training the model on a large-scale sketch dataset and fine-tuning it on a hand-drawn HTP sketch dataset.
• Cross-validation was utilized for robust performance evaluation on the HTP dataset.

Main Results:

• FBANet achieved a maximum accuracy of 99.07% and an average accuracy of 97.71% on the HTP sketch dataset.
• The proposed one-stage approach demonstrated superior performance compared to traditional classification models and prior works.
• The model exhibited a simple data preprocessing pipeline and calculation process, indicating high automation.

Conclusions:

• The FBANet model, after pre-training, shows exceptional performance for depression recognition in HTP sketches.
• This deep learning approach offers a promising, automated, and accurate tool for the auxiliary diagnosis of depression.
• FBANet addresses the limitations of subjectivity and low automation in traditional HTP sketch analysis for mental health assessment.