BLAN: Bi-directional ladder attentive network for facial attribute prediction

Highlights：

• A novel Bi-directional Ladder Attentive Network (BLAN) to make facial attribute prediction better.

• Learning hierarchical representations for exploiting the correlations between feature hierarchies and attribute characteristics.

• Residual Dual Attention Module (RDAM) shows the excellent ability in interweaving features from the encoder and the decoder.

• Local Mutual Information Maximization (LMIM) loss further incorporates the locality of the input attribute features to the high-level representations and produces high-quality features.

• Adaptive score fusion module performs well in merging multiple global and local decisions from all hierarchies.

摘要

•A novel Bi-directional Ladder Attentive Network (BLAN) to make facial attribute prediction better.•Learning hierarchical representations for exploiting the correlations between feature hierarchies and attribute characteristics.•Residual Dual Attention Module (RDAM) shows the excellent ability in interweaving features from the encoder and the decoder.•Local Mutual Information Maximization (LMIM) loss further incorporates the locality of the input attribute features to the high-level representations and produces high-quality features.•Adaptive score fusion module performs well in merging multiple global and local decisions from all hierarchies.

论文关键词：Deep facial attribute prediction,Bi-directional ladder attentive network (BLAN),Residual dual attention module (RDAM),Local mutual information maximization (LMIM),Adaptive score fusion

论文评审过程：Received 13 March 2019, Revised 22 October 2019, Accepted 8 December 2019, Available online 10 December 2019, Version of Record 15 December 2019.