MPDRM: A Multi-Scale Personalized Depression Recognition Model via facial movements

Automatic depression recognition based on facial movements in videos has become a research hotspot. However, existing methods tend to confuse individual inherent facial behavioral habits with characteristics specific to depression, which leads to misjudgments. To address this, we propose a Multi-scale Personalized Depression Recognition Model (MPDRM) that mitigates the negative impact of individual differences, enabling the model to focus on general and robust facial depression cues. The proposed model consists of three main components: the Multi-scale Depression Feature Network (MDFN), the Multi-scale Personality Feature Network (MPFN), and the Relational Attention Recognition Module (RARM). The MDFN extracts depression-related information, while the contrastive learning-based MPFN extracts stable personalized information. In both MDFN and MPFN, we insert the Multi-scale Motion Pattern Extraction Module (MMP) to capture rich multi-scale spatiotemporal facial features. Finally, the RARM is designed to enhance the representation of depression and output the results. Cross-validation on a specifically constructed longitudinal dataset demonstrates that our model outperforms other models. Experimental results indicate that suppressing personalized information of facial movements can effectively improve the accuracy of depression recognition.