MTDAN: A Lightweight Multi-Scale Temporal Difference Attention Networks for Automated Video Depression Detection