Learning to Hash with Binary Deep Neural Network

Abstract

This work proposes deep network models and learning algorithms for unsupervised and supervised binary hashing. Our novel network design constrains one hidden layer to directly output the binary codes. This addresses a challenging issue in some previous works: optimizing non-smooth objective functions due to binarization. Moreover, we incorporate independence and balance properties in the direct and strict forms in the learning. Furthermore, we include similarity preserving property in our objective function. Our resulting optimization with these binary, independence, and balance constraints is difficult to solve. We propose to attack it with alternating optimization and careful relaxation. Experimental results on three benchmark datasets show that our proposed methods compare favorably with the state of the art.

Unsupervised Deep Hashing With Pseudo Labels for Scalable Image Retrieval

“…On the basis of DH, Deepbit adds another constraint, rotation invariant, to the final loss function, and achieve a slight boost on the performance. Do et al [41] propose a Binary Deep Neural Network (BDNN), which has both unsupervised and supervised versions. BDNN introduces one hidden layer to directly output the binary codes and then utilises the binary code layer to reconstruct the original feature.…”

“…We systematically compare our method with eight state-of-the-art non-deep methods: ITQ [22], PCAH [20], LSH [13], DSH [55], SpH [44], SH [19], AGH [21], and SELVE [56], and three deep methods: DH [39], Deepbit [40], UH-BDNN [41] for retrieval task, all these eleven methods are unsupervised. All of the non-deep methods and UH-BDNN [41] in our experiments use the same VGG [58] fc7 feature as that in our method, and DH [39] and Deepbit [40] are based the same settings like that in their original papers.…”

“…Fortunately, at near recent, on account of great development in the area of deep learning [31]- [34], hashing methods turn to extract deep features automatically using Convolutional Neural Network (CNN) descriptors [35]- [38], which inherit high discriminative property and often get state-of-the-art performance. However, most state-of-the-art unsupervised deep hashing algorithms [39]- [41] suffer from severe performance degradation due to lack of label information, e.g. DH [39] and Deepbit [40] use quantisation loss and evenly distribution loss, which can only make the binary codes have large variance, but can not get the real distribution of the data, UH-BDNN [41] adds a reconstruction layer after the binary code layer, which can obtain the semantics of the images, but still can not get the category information, which is more important than the semantics for image retrieval, and the experimental results reported in their original literatures and this paper can also show this point.…”

“…However, most state-of-the-art unsupervised deep hashing algorithms [39]- [41] suffer from severe performance degradation due to lack of label information, e.g. DH [39] and Deepbit [40] use quantisation loss and evenly distribution loss, which can only make the binary codes have large variance, but can not get the real distribution of the data, UH-BDNN [41] adds a reconstruction layer after the binary code layer, which can obtain the semantics of the images, but still can not get the category information, which is more important than the semantics for image retrieval, and the experimental results reported in their original literatures and this paper can also show this point. Therefore, it is necessary to find out the visual feature distribution for unsupervised hashing.…”

Compact Hash Codes for Efficient Visual Descriptors Retrieval in Large Scale Databases

“…Do et al [25] have addressed the problem of learning binary hash codes for large scale image search using a deep model which tries to preserve similarity, balance and independence of images. Two sub-optimizations during the learning process allow to efficiently solve binary constraints.…”