A multi-scale transformer framework with consistency and dual-view for depression detection

Automatic depression detection (ADD) from multimodal interviews offers unobtrusive mental-health screening, yet current systems often miss faint behavioral cues, suffer under noisy inputs, and rely heavily on annotated data. We propose MSCDV, a transformer-based framework where three tightly integrated mechanisms form a unified learning loop. A multi-scale encoder captures local facial, vocal, and lexical nuances while maintaining global context; a perturbation-based consistency loss enforces alignment between clean and noised views, improving robustness; and a dual-view stepwise schedule presents complementary modality subsets in alternation, enhancing generalization under data scarcity. Because all components are co-optimized, improvements in one propagate through the system, producing features that are fine-grained, noise-aware, and data-efficient. On the DAIC-WOZ and E-DAIC benchmarks, MSCDV achieves mean F1-scores of 0.87 and 0.82, respectively, outperforming a wide range of state-of-the-art baselines without requiring additional supervision. Extensive ablations confirm that removing any single component results in performance degradation, underscoring their synergistic contribution to robust, real-world ADD.