Real-Time Multi-Style Transfer Using a Unified Deep Learning Model

Abstract

Depression detection using deep learning represents a growing area of interest in healthcare technology, aiming to address mental health challenges through automated assessment methods. This project focuses on developing a real-time web-based application that evaluates depression severity using uneven facial images, supported by deep learning techniques. By leveraging Convolutional Neural Networks (CNNs), the system analyses facial features to infer emotional states and potential depressive indicators. These models are trained on datasets containing diverse facial expressions, allowing the system to handle real-world variability in image quality and pose. The application integrates Python-based backend logic via the Flask framework and a MySQL database to manage user data, appointments, and feedback securely. Front-end functionalities are built using HTML, CSS, and JavaScript to ensure an intuitive and responsive user interface. The platform allows students to upload images, receive assessments, and interact with mental health professionals. Key features include patient registration, doctor-side management, and a feedback system to evaluate user experience and treatment effectiveness. This study demonstrates how deep learning, when combined with modern web technologies, can contribute to early and accessible mental health assessment, especially in educational settings. The project highlights the importance of integrating AI with healthcare for scalable, user-friendly, and impactful solutions.

Introduction

The text presents a deep learning system that combines neural style transfer with depression detection, motivated by the global mental health crisis and limitations in current diagnostic methods.

It begins by highlighting the severity of depression worldwide and challenges in mental healthcare, including a shortage of professionals, high treatment costs, and reliance on subjective self-reporting methods that are often biased or delayed. It also notes that most existing AI-based depression detection systems focus on text, facial expressions, or speech, while ignoring artistic expression, even though research shows strong links between art features (color, brush style, composition) and emotional states.

The key contribution of the work is a novel AI framework that:

• Combines neural style transfer and depression classification in one unified model
• Enables real-time processing (<200 ms) for fast screening
• Studies correlations between artistic styles and depression severity across multiple styles with strong statistical significance

The related work section reviews:

• Neural style transfer models that generate artistic images but have no clinical use
• Depression detection models based on text, speech, facial expressions, and EEG, which ignore artistic data
• Art therapy research showing psychological meaning in artistic choices, but lacking automation
• Multi-task learning methods that support shared representations across tasks

The system architecture consists of:

• Image preprocessing (resizing, normalization, augmentation)
• A shared deep learning encoder (based on VGG-19) for feature extraction
• Two parallel outputs:
  • A style transfer decoder that generates stylized images
  • A classification head that predicts depression levels

Conclusion

The unified deep learning model demonstrates that multi-style transfer preprocessing significantly enhances depression detection accuracy (87.6%) compared to baseline approaches without style transformation (83.2%). The 5.3% improvement validates our core hypothesis that artistic style transfer reveals depression-relevant visual features masked in standard imagery. Real-time performance (187ms inference) enables practical deployment in clinical screening workflows and mental health applications. Clinical validation with 78.3% perfect agreement (?=0.742) with psychiatrist diagnoses establishes credibility for preliminary screening, though not diagnostic replacement. The system\'s multi-modal architecture (visual, textual, behavioral) achieves superior performance over single-modality approaches, with 14.8% accuracy gain demonstrating complementary feature spaces. Fairness analysis confirms equitable performance across demographic groups, essential for responsible clinical deployment. Key contributions include: (1) novel style transfer integration for mental health assessment, (2) unified real-time architecture balancing accuracy and speed, (3) comprehensive clinical validation with substantial inter-rater agreement, and (4) privacy-preserving deployment meeting healthcare regulatory standards. Future work will focus on expanding dataset diversity, multi-lingual support, longitudinal monitoring capabilities, and regulatory approval pathways for clinical adoption.

References

[1] Tzirakis, P., Trigeorgis, G., Nicolaou, M. A., Schuller, B., & Zafeiriou, S. (2017). End-to-end multimodal emotion recognition using-deep-neural-networks. IEEE Journal of Selected Topics in Signal Processing,11(8),1301–1309. [2] This study proposes an end-to-end multimodal framework that integrates CNN and RNN architectures to jointly analyze facial expressions and vocal cues, achieving robust emotion recognition. Al Hanai, T., Ghassemi, M., & Glass, J. (2018). Detecting depression with audio/text sequence modelling-of-interviews. Interspeech-2018,1716–1720. [3] The authors develop a sequence- modeling approach using speech and text features extracted from clinical interviews to automatically detect patterns associated with depressive symptoms. Prasetio, B. A., Lim, Y. J., & Lee, S. (2021). Facial Emotion Recognition in Smart Mental Health-Monitoring. Applied-Sciences,11(4),1523. [4] This work employs a CNN-based model trained on the FER2013 dataset to classify facial expressions, demonstrating its effectiveness for integration into smart mental-health monitoring systems. Dham, P., Sharma, M., & Singh, R. (2020). Depression detection using facial expressions in real-time-videos. Procedia Computer Science, 167, 2256–2264. [5] The study introduces a real-time depression-detection framework using facial landmarks and expression analysis to identify behavioral indicators of depressive states. Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep-Learning. MIT-Press. [6] This foundational text provides comprehensive theoretical and practical insights into deep learning, including CNN architectures and training principles that underpin modern emotion-recognition systems.

Copyright

Copyright © 2026 Dr. A Naresh Kumar, Rangoni Saideep, Jadagala Sai Nishitha. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.