Channel-Dcoupled and Local-Attention Network for Depression Detection

To address the problems of high redundancy and lack of attention to local features, this paper proposes a Channel-decoupled and Local-attention Network (CDLA-Net) for depression detection. The proposed network consists of Channel-decoupled subnet (CDS) and Local-attention subnet (LAS). In the CDS, to reduce the redundancy caused by repeated extraction of channel features, a 3D Inverted Residual (3D-IR) structure is constructed to learn fine-grained semantics related to depression through independent convolutions in the channel dimension. In the LAS, to avoid the loss of details caused by insufficient attention to local features, graph convolution operation is embedded to focus on local details that are highly correlated with depression. The experimental results on the AVEC2014 dataset show that the MAE and RMSE of the proposed method reach 6.03 and 7.60 respectively, which are 0.69 and 1.33 lower than those of the baseline network. Compared with other related methods, the proposed end-to-end method, which effectively reduces the redundant features and focuses on detail semantics, improves the performance of depression detection.