Research map / Label-efficient and robust learning / Recognition and long-tail
OpticalDR: A Deep Optical Imaging Model for Privacy-Protective Depression Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes OpticalDR, a learnable thin-lens optical front end jointly trained with progressive face-analysis tasks (identity, emotion, depression) to record de-identified face images for privacy-protective depression score regression.
Problem
Automatic depression recognition from facial video requires facial images that expose patient identity, creating privacy risks; existing privacy approaches (landmarks, blurring) still need visible faces at capture or inference.
Approach
Parameterize a single thin lens (Zernike coefficients, wave-based PSF) as a learnable optical layer, co-optimized end-to-end with a SANet (ResNet + spatial attention) trained in four progressive steps (identity obfuscation with inverse triplet and landmark losses, emotion classification, depression histogram loss, and a fused MSE score head) so the sensor captures only identity-erased, depression-
Key results
- OpticalDR reaches MAE/RMSE 7.53/8.48 on AVEC 2013 and 7.89/8.82 on AVEC 2014.
- Privacy: average AUC of 0.51 across FaceNet, IR152, IRSE50, ArcFace on CelebA, versus 0.91 with no privacy processing.
- Beats ViT + lens and STA-DRN (SA) + lens but trails the non-private LQGDNet (6.38/8.20 on AVEC 2013).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add emotion-task pretraining before depression training (ablation) | OpticalDR · AVEC 2013 | MAE: 10.35 -> 9.18 | Helped |
| Add emotion-task pretraining before depression training (ablation) | OpticalDR · AVEC 2014 | MAE: 10.31 -> 9.06 | Helped |
| Add identity-lens pretraining before depression training (ablation) | OpticalDR · AVEC 2013 | MAE: 10.35 -> 10.15 | Helped |
| Add identity-lens and emotion-task pretraining before depression training (ablation) | OpticalDR · AVEC 2013 | MAE: 10.35 -> 8.54 | Helped |
| Add identity-lens and emotion-task pretraining before depression training (ablation) | OpticalDR · AVEC 2014 | MAE: 10.31 -> 8.87 | Helped |
| Add auxiliary identity, emotion, and depression pretraining (ablation) | OpticalDR · AVEC 2013 | MAE: 8.46 -> 7.92 | Helped |
| Add a privacy lens to ViT | ViT · AVEC 2013 | RMSE: 10.67 -> 10.49 | Helped |
| Add a privacy lens to LQGDNet | LQGDNet · AVEC 2013 | MAE: 6.38 -> 10.56 | Hurt |
| Add a privacy lens to STA-DRN (SA) | STA-DRN (SA) · AVEC 2014 | MAE: 7.27 -> 8.67 | Hurt |
| Use OpticalDR instead of LQGDNet | Depression recognition models · AVEC 2013 | MAE: 6.38 -> 7.53 | Hurt |
| Use OpticalDR instead of LQGDNet | Depression recognition models · AVEC 2014 | MAE: 6.08 -> 7.89 | Hurt |
Surprising findings
- Simply adding an optical lens to existing DR models (ViT, STA-DRN) does not notably improve their depression error, and LQGDNet + lens degrades markedly (10.56 MAE on AVEC 2013).
Limitations
- Falls short of the non-private LQGDNet on DR error, which uses visible facial landmarks/quaternion features.
- Evaluated only in simulation; physical deployment of the optical system is left to future work.
- Only single-frame image input; temporal information not used.
Experimental setup
Datasets: CelebA CK+ AVEC 2013 AVEC 2014
Benchmarks: AVEC 2013 AVEC 2014 CelebA
Metrics: MAE RMSE AUC
Baselines: ViT LQGDNet STA-DRN (SA) Gaussian blur defocus blur ViT + lens STA-DRN (SA) + lens
Builds on: Deep Optics ResNet STA-DRN LQGDNet histogram loss
Closest papers by meaning
- How Private is Private? A Comparative Study for Face De-Identification (NeurIPS 2026 · similarity 0.85)
- Privacy-Preserving Optics for Enhancing Protection in Face De-Identification (CVPR 2024 · similarity 0.84)
- Validating Privacy-Preserving Face Recognition under a Minimum Assumption (CVPR 2024 · similarity 0.82)
- Breaking Data Silos in Parkinson’s Disease Diagnosis: An Adaptive Federated Learning Approach for Privacy-Preserving Facial Expression Analysis (AAAI 2025 · similarity 0.81)
- PID: Prompt-Independent Data Protection Against Latent Diffusion Models (ICML 2024 · similarity 0.81)
- PerceptAnon: Exploring the Human Perception of Image Anonymization Beyond Pseudonymization for GDPR (ICML 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.