点击次数：

论文类型：期刊论文

发表时间：2019-01-01

发表刊物：OPTIK

收录刊物：SCIE、EI

卷号：180

页面范围：847-857

ISSN号：0030-4026

关键字：High-order statistics; Supervised hashing; Bilinear pooling; CNN; Image retrieval

摘要：Recently, deep hashing has achieved excellent performance in large-scale image retrieval by simultaneously learning deep features and hash function. However, state-of-the-art methods for this task have so far failed to explore feature statistics higher than first-order. To address this problem, we propose two novel deep high-order supervised hashing (DHoSH) architectures based on point-wise labels (DHoSH-PO) and pair-wise labels (DHoSH-PA), respectively. The core of DHoSH is a trainable layer of bilinear features incorporated into existing deep convolutional neural network (CNN) for end-to-end learning. This layer captures the interaction of local features by bilinear pooling, using correlation to model dependencies of features within the same layer or cross-correlation for features across different layers of the CNN. Furthermore, DHoSH systematically employs the high-order statistics of features of multiple layers. Extensive experiments on commonly used image retrieval benchmarks show that our DHoSH-PO and DHoSH-PA models yield improved accuracy over its first-order counterparts and achieve effective performance for these benchmarks.