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

摘要: Abstract Deep facial attribute prediction has received considerable attention with a wide range of real-world applications in the past few years. Existing works almost extract abstract global features at high levels deep neural networks to make predictions. However, local low levels, which contain detailed information, are not well exploited. In this paper, we propose novel Bi-directional Ladder Attentive Network (BLAN) learn hierarchical representations, covering correlations between feature hierarchies and characteristics. BLAN adopts layer-wise bi-directional connections based on autoencoder framework from levels. way, characteristics could be correspondingly interweaved each level via multiple designed Residual Dual Attention Modules (RDAMs). Besides, derive Local Mutual Information Maximization (LMIM) loss further incorporate locality attributes high-level representations hierarchy. Multiple classifiers receive produce decisions, followed by proposed adaptive score fusion module merge these decisions for yielding final result. Extensive experiments two datasets, CelebA LFWA, demonstrate that our outperforms state-of-the-art methods.