Figure 1:

Overview of the proposed network architecture. It has three branches: two channel-specific networks (FiA-Net_con and FiA-Net_mor, top and bottom paths in blue colors) provide channel-specific age estimation results, and one attention-driven fusion network (FiA-Net_fus, middle path in orange) provides the final age estimation results. In channel-specific branches, the Res_i(i = 1, 2, 3, 4) boxes are residual network blocks, and $f_i^{m_i}$ are the intermediate deep features of channel image m_i after the i^th residual block. In the fusion branch, the F_i boxes are the fusion blocks. GAP is the global average pooling.