点击次数：

论文类型：期刊论文

发表时间：2020-04-01

发表刊物：PATTERN RECOGNITION

收录刊物：EI、SCIE

卷号：100

ISSN号：0031-3203

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

摘要：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 of deep neural networks to make predictions. However, local features at low levels, which contain detailed local attribute information, are not well exploited. In this paper, we propose a novel Bi-directional Ladder Attentive Network (BLAN) to learn hierarchical representations, covering the correlations between feature hierarchies and attribute characteristics. BLAN adopts layer-wise bi-directional connections based on the autoencoder framework from low to high levels. In this way, hierarchical features with local and global attribute characteristics could be correspondingly interweaved at each level via multiple designed Residual Dual Attention Modules (RDAMs). Besides, we derive a Local Mutual Information Maximization (LMIM) loss to further incorporate the locality of facial attributes to high-level representations at each hierarchy. Multiple attribute classifiers receive hierarchical representations to produce local and global decisions, followed by a proposed adaptive score fusion module to merge these decisions for yielding the final prediction result. Extensive experiments on two facial attribute datasets, CelebA and LFWA, demonstrate that our BLAN outperforms state-of-the-art methods. (C) 2019 Elsevier Ltd. All rights reserved.