Dynamic Multi-View Hashing for Online Image Retrieval

Transcription

1 Proceedings of the wenty-sixth International Joint Conference on Artificial Intelligence IJCAI-17 Dynaic Multi-View Hashing for Online Iage Retrieval Liang Xie 1, Jialie Shen 2, Jungong Han 3, Lei Zhu 4, Ling Shao 5 1 Wuhan University of echnology, China 2 Northubria University, United Kingdo 3 Lancaster University, United Kingdo 4 he University of Queensland, Australia 5 University of East Anglia, United Kingdo whutxl@hotail.co, {jialie, jungonghan77, leizhu0608}@gail.co, ling.shao@ieee.org Abstract Advanced hashing technique is essential in large s- cale online iage retrieval and organization, where iage contents are frequently changed. While traditional ulti-view hashing ethod has achieve proising effectiveness, its batch-based learning based schee largely leads to very expensive updating cost. Meanwhile, existing online hashing schee generally focuses on single-view data. Good effectiveness can not be expected when searching over real online iages, which typically have ultiple views. Further, both types of hashing ethods only can generate hash codes with fixed length. hus they have liited capability on coprehensive characterization of streaing iage data. In this paper, we propose dynaic ulti-view hashing, which can adaptively augent hash codes according to dynaic changes of iage. Meanwhile, leverages online learning to generate hash codes. It can increase the code length when current code is not able to represent new iages effectively. Moreover, to gain further iproveent on overall perforance, each view is assigned with a weight, which can be efficiently updated in the online learning process. In order to avoid the frequent updating of code length and view weights, an intelligent buffering schee is designed to preserve significant data to aintain good effectiveness of. Experiental results on t- wo real-world iage datasets deonstrate superior perforance of DWVH over several state-of-theart hashing ethods. 1 Introduction With the explosive growth of online iage repositories, techniques to facilitate fast and effective large scale content access are getting ore and ore attentions. In recent years, hashing has eerged as a proising technology to support large-scale iage retrieval [Mu et al., 2010; Zhen and Yeung, 2012a; Guo et al., 2017b; Wu et al., 2017; Corrsponding Author Zhu et al., 2017]. Except various visual features, the Web iages are associated with various text inforation such as tags or short coents. In order to iprove Web iage search perforance, various ulti-view hashing ethods have been recently proposed [Zhang et al., 2011; Wu et al., 2014; 2015; Xie et al., 2016; Liu et al., 2017; Lin et al., 2017b; 2017a; Guo et al., 2017a]. Multi-view hashing integrates the advantages of hashing technology and ulti-view learning, it enjoys better perforance than single-view hashing, and can support the queries fro different views such as visual features or text features. Web iage database is highly dynaic and frequently updated. raditional ulti-view hashing ethods are developed based on batch-based learning, which is very expensive in ters of updating cost. Multi-view hashing ethods ay accuulate all data to retrain the hash codes and functions when new iages arrive, however their tie coplexity will be extreely high and unaffordable. Recently, several online hashing ethods [Leng et al., 2015; Cakir and Sclaroff, 2015b] are proposed to support effective search over dynaic iage databases. However, their perforance is far beyond satisfactory. When applying existing ulti-view hashing and online hashing ethods for web iage retrieval, they generally suffer fro several issues. Existing online hashing ethods ai at generating the binary codes with fixed length, which is not suited to the dynaic iage data. In real world, the contents of new iage data could be seantically different fro the old one. When new iage data is inserted into the database, longer codes should be required to preserve ore discriinative inforation to support search over updated iage database. According to previous results [Zhou et al., 2014], longer hash code can lead to ore coprehensive iage representation and better search perforance. For exaple, Wikipedia dataset [Rasiwasia et al., 2010] consisting of 2866 iages only needs 32 bits codes to achieve the best perforance. However, for NUS-WIDE having iages, at least 128 bits are required to achieve the best perforance. However, larger code length does not ean the better perforance, in that hashing for data with sall size ay get trapped in local inia [Zhen and Yeung, 2012b]. herefore, general hashing ethods usually choose an appropriate code length for a particular database, which is not suitable to dynaic database, whose size is continuously changed. o solve this proble, the code length for dynaic iage 3133

2 Proceedings of the wenty-sixth International Joint Conference on Artificial Intelligence IJCAI-17 database needs to be intelligently augented to guarantee hashing perforance. Another proble is how to easure iportance of different views in the online learning of hash codes. Although each view will contribute to the final hash codes, they can not be equally treated. For exaple, text feature usually contains ore seantic inforation than visual feature [Caicedo et al., 2012], thus it should be assigned with the higher weight. Several ulti-view hashing ethods take the view weights into account and thus they achieve significant perforance iproveent. However, to the best of our knowledge, none of existing online hashing ethods focus on developing intelligent schee to estiate optial weighs for different views. In this paper, we propose a novel unsupervised online hashing ethod: Dynaic Multi-view Hashing to support the effective and efficient dynaic online iage retrieval. owards this goal, uses a dynaic hashing schee, where length of hashing codes can be adaptively augented. In, a dictionary is used to construct the hash codes of iages. When a new data cannot be represented by current dictionary, it will be added to augent the dictionary. Moreover, applies ulti-view features for hashing, and proper weight is assigned to each view. In the online learning process, to avoid the frequent updating of dictionary and odality weights, a buffer is used to preserve significant iages for purpose of subsequential updating. he core technical contributions of our work can be suarized as follows: In, the code length can be dynaically augented with the changes of iage data. As a result, can better represent the streaing iages. In addition, users are not required to predefine the appropriate code length which will be autoatically obtained by. can effectively estiate proper weights for different views to reflect their different iportance in hashing. Previous online hashing ethods generally don t have weighting schee. An intelligent buffer schee is designed for the online learning process of, which enables highly efficient learning process of hash codes and view weights. Experiental results on real-world iage datasets show the superiority of in ters of both search efficiency and accuracy. he reainder of this paper is organized as follows: Section 2 introduces related work; Section 3 coprehensively introduces the proposed ulti-view hashing ethod; Section 4 introduces the experiental configurations and reports the experiental results and ain findings. Finally, Section 5 concludes the paper with future work. 2 Related Work Multi-view Hashing - In recent years, uch attention has been paid for ulti-view hashing. Most ulti-view hashing ethods can be divided to two categories according to the query types supported. Several ulti-view hashing ethods cobine all features to construct database hash codes, but they require that query has the sae features as database Database Dynaic Hash Codes Buffer Iage Hash Function ext Hash Funciton Multi-view Dictionary Learning Figure 1: he overall illustration of. Visual Query ext Query New Iage data. his type of ulti-view hashing usually ais to gain the effective cobination of ulti-view features. Coposite Hashing with Multiple Inforation Sources CHMIS [Zhang et al., 2011] uses graphs of ulti-views features to learn the hash codes, and each view is assigned with a weight for cobination. Since graph learning is inefficient, anchor graph is adopted to accelerate the learning process. Multi-view Anchor Graph Hashing MVAGH [Ki and Choi, 2013] exploits anchors to construct graphs for hashing. Multiview Alignent Hashing MAH [Liu et al., 2015] cobines atrix factorization with anchor graph learning for hashing, and it achieves good perforance based on ulti-view features. Multi-view ethods also can support queries based on different views. his type of ulti-view hashing is also called as cross-view hashing. Generally, cross-view hashing ethods ai at odeling the correlation of different views. Crossview Hashing CVH [Kuar and Udupa, 2011] optiizes the haing distance of different views, it can be regarded as an extended version of Canonical Correlation Analysis C- CA. Inter-edia Hashing [Song et al., 2013] both optiizes the intra-edia and inter-edia consistency. Seantic Correlation Maxiization SCM [Zhang and Li, 2014] axiizes the seantic correlation between iage and text features, and it further considers the quantization loss of hash codes. Seantic opic Multiodal Hashing SMH [Zhou et al., 2014] firstly generates text topics and iage concepts, and then correlates the in a coon hash space. Online Hashing is ore practical than traditional batchbased learning ethods, in that any new iages in realworld are generated. In recent three years, a few online hashing ethods are proposed. Online Kernel Hashing OKH [Huang et al., 2013] and Adaptive Hashing [Cakir and Sclaroff, 2015a] are supervised ethods, they use siilar iage pairs as training data. Although they can effectively p- reserve the correlation of iages in hash codes, they require labeled training exaples which are difficult to collect. Online Sketching Hashing OSH [Leng et al., 2015] is unsupervised online hashing ethod, thus it is not relied on labeled data. he ain disadvantage of existing online hashing ethods is that they are not designed for ulti-view features, so they cannot perfor well for the retrieval of ulti-view iages. 3 Dynaic Multi-View Hashing consists of two ain odules: online learning process and search process. Figure 1 illustrates the overall structure of. In the dynaic hash codes, blue area, red area 3134

3 Proceedings of the wenty-sixth International Joint Conference on Artificial Intelligence IJCAI-17 and yellow area denote the old codes, hash codes of new data and the augented codes respectively. he database consists of any iages having both visual and text contents. When new iage arrives, ultiple kinds of visual features and one text feature are extracted and cobined via ulti-view dictionary learning. he differentiation threshold L t deterines whether the iage can be represented by current hash codes. If not, it is added to the buffer, and the buffer data are used to augent hash codes till no ore space is available in the buffer. In the following sections, ore details about L t will be presented. 3.1 Model Forulation In, at step t, the database contains old iage features added at previous steps Xt 1 M, where M is the nuber of views. Xt 1 R t 1 d, and d is the diension of th view feature. he old iages are represented with hash codes H t 1 {0, 1} t 1 C, where C is the code length. hen new iage x t M is added into the database, and Xt = [Xt 1, x t ]. Our goal is to learn new hash codes h t {0, 1} 1 C to represent new iage. he hash function also needs to be learnt to project any views into the learned hash space. owards the goal, the kernel dictionary representation is gained to construct hash codes. Dictionary learning is effective in learning hash codes [Yu et al., 2014; Zhang et al., ], and the kernel projection will ake the hash codes preserve ore discriinative inforation. hus, at tie t, the dictionary consists of a subset of the database saples with uti-view features. For each view, its corresponding dictionary is denoted as Dt = [x a 1,..., x a C ]. Assuing each saple can be constructed fro its hash codes and dictionary, we have the following forulation: in L t Ht, Dt M = α 2 H t ϕ Dt ϕ Xt 2 F + λ H t 2 F s.t. i=1 α = 1 1 where ϕ is a kernel projection function, λ is the regularization paraeter, α = [α 1,..., α M ] denotes the view weights. 3.2 Online Learning Basic process At each step t, when new iage is added into the database, the overall objective function can be reforulated as: in L t Ht, Dt M = L t 1 Ht 1, Dt M + lt ht, Dt M If dictionary Dt M is not changed, Dt = Dt 1, the constraint L t 1 Ht 1, Dt M has been satisfied by previous learning steps. hus, We only need to consider the optiization of l t ht, Dt M, and the objective function is: in l t ht, Dt M = α 2 h t ϕ Dt ϕ x t 2 F + λ h t 2 F 3 2 After applying the kernel trick, the objective function 3 can be rewritten as: α 2 h t K t h t 2h t zt + ktt + λh t h t 4 where K t R C C is the kernel atrix of the dictionary D t, and K t i, j = ϕx a i ϕx a j, z t i = ϕx t ϕx a i, a i is the index of i th dictionary eleent. k tt = ϕx t ϕx t. By setting the derivative of Eq.4 to zeros, we can obtain h t as: where K t = M h t = z t Kt + λi 1 5 α 2 K t, z t = M α 2 z t. For each odality, if vector ϕ x t is linearly independent on vectors of dictionary ϕ Dt. hen h t coputed by Eq.5 cannot satisfy Eq.3, and x t should be added into the dictionary. herefore, we have to deterine the linear dependence of new vectors. By substituting Eq.5 into Eq.3, it can be transfored to: l t ht, Dt 1 M M = α 2 ktt zt h t + λh t h t 6 We use a threshold ρ to deterine the linear dependence. If l t ht, D t 1 M < ρ, then the approxiate linear dependence described by Eq.3 can be satisfied. And we do not need to change dictionary. If l t ht, D t 1 M ρ, then new hash codes cannot be effectively represented by current dictionary. As a result, both dictionary and hash codes need to be augented, and the odal weights should be recoputed accordingly. Dictionary augent with buffer schee Considering the worst case that new iage is always very d- ifferent to database iages and cannot be represented by current dictionary, we have to frequently update the dictionary, which is obviously inefficient. herefore, a buffer schee is applied to avoid the frequent updating process. he buffer schee relaxes the restrict linear dependence in Eq.3. At each step t, if the new iage satisfies l t ht, Dt 1 M ρ, then iage t is added into the buffer with axiu size B ax. After the updating, in case that the buffer is full, iage b will be selected fro the buffer with highest value of l t. hen we augent the dictionary with iage b and update the view weights α. he dictionary is updated with Dt = [Dt 1, x b ], and the hash codes constructed fro dictionary have to be augented accordingly. We should optiize the following overall objective function: in L t Ht, Dt M = L t\b Ht\b, Dt M + lb hb, Dt M s.t. α =

4 Proceedings of the wenty-sixth International Joint Conference on Artificial Intelligence IJCAI-17 Where H t\b denotes the hash codes of database iages except b. Since iage b is linearly independent on other dictionary iages, we can only use it to represent itself. herefore we have h b = [0, 1], and l b hb, D t M = 0. hen Eq.7 can be siplified as: + M L t Ht, Dt M = λ r Ht Ht H t\b K t Ht\b 2H t\b α 2 r Z t\b + K t\b 8 After the dictionary is augented, the new K t and Z t\b can be coputed as: [ ] K t = K t 1 zb z b k bb [ Zt\b = Zt 1\b, k ] X t\b, x b 9 Lea 1. If H t\b = [ H t 1\b, 0 ], then the following equation is satisfied: L t Ht\b, Dt M = Lt 1 Ht 1\b, Dt 1 M 10 Proof. Fro Eq.9, we can obtain: H t\b K t Ht\b = [ H t 1\b, 0 ] [ K t 1 zb z b k bb Siilarly, we have: ] [ H t 1\b 0 = H t 1\b K t 1 H t 1\b 11 H t\b Z t\b = Ht 1\b Z t 1\b 12 Based on above two equations and Eq.8, we can easily obtain that each eleent in L t and L t 1 is equivalent, thus Eq.10 is satisfied. heore 1. he objective function [ L t Ht, ] Dt M will be decreased after we set H t = Ht\b, h b, and H t\b = [ Ht 1\b, 0 ] and h b = [0, 1]. Proof. According to Lea 1 and Eq.7, we can obtain that L t Ht, D t M = Lt 1 Ht 1\b, D t 1 M + l b hb, D t M, and lb hb, D t M = 0 can be satisfied by h b = [0, 1]. Before the dictionary augent, iage b in the buffer always has a high l b hb, D t 1 M which describes the linear dependence, where h b denote old codes of b. It is obviously that l b hb, D t 1 M > ρ > 0 = l b hb, D t M. Finally we have Lt Ht, D t M < L t 1 Ht 1, D t 1 M. herefore, the objective function can be decreased. heore 1 confirs that our objective function is consistently decreased by the buffer schee. By specially considering the iage with largest l b and reduce it to 0 by using dictionary augent, the objective functions L t can be largely decreased. If the buffer is full and we augent the codes, then we preserve half iages with highest l to ensure that ] Algorith 1 Online learning process of at step t. Input: x t M,Dt 1 M, K t 1 M, α Output: H t, Dt M, K t M, α 1: Copute h t by Eq.5; 2: Copute l t h t, Dt 1 M by Eq.6 3: if l t h t, Dt 1 M < δ then 4: H t = [Ht 1, sgnh t ] ; 5: K t = K t 1 and D t = D t 1; 6: else if l t h t, D t 1 M >= δ then 7: Add t into the buffer, and use Algorith 2 to optiize H t, α and K t ; 8: end if Algorith 2 Augent dictionary with buffer schee. Input: Buffer, H t, Dt 1 M, K t 1 M, α Output: H t, Dt M, K t M, α 1: Add iage t and corresponding l t h t, Dt 1 M into buffer; 2: Buffer size B = B + 1; 3: if B = B ax then 4: Select iage b in buffer with highest l b ; 5: Reove 1/2 of buffer data with lowest l; 6: Augent dictionary Dt = [Dt 1, x b ] ; 7: Update K t according to Eq.9; 8: Update H t\b = [ H t 1\b, 0 ] ; 9: Update h b = [0, 1]; 10: Update α according to Eq.13 11: else if B < B ax then 12: K t = K t 1, Dt = Dt 1; 13: end if high l can be preserved for further process to decrease objective function. he frequency of augenting depends on the buffer size, if buffer size is large, then the learning process is ore efficient, but objective function can not be effectively decreased. hus we should ake a tradeoff between learning efficiency and effect by adjusting the buffer size. At last, the view weights need to be updated to further reduce the objective function 7. by using the lagrange ultiplier for Eq.7, we can copute α by: where δ = r α = 1/δ M i=1 1/δ i 13 H t K t Ht 2H t Zt + Ktt he basic online learning process and dictionary augentation process are shown in Algorith 1 and 2 respectively. is efficient in learning hash codes. Fro Algorith 1 we can find that the only the hash codes of new data is coputed, and the old codes are not needed to update. Although Algorith 2 updates the old hash codes of previous. 3136

5 Proceedings of the wenty-sixth International Joint Conference on Artificial Intelligence IJCAI MAH OSH x 10 4 x 10 4 a Iage query b ext query Figure 2: he scores of MIR Flickr at different database sizes x 10 4 c Multi-view query MAH OSH x 10 5 x 10 5 a Iage query b ext query Figure 3: he scores of NUS-WIDE at different database sizes x 10 5 c Multi-view query data, they are only augented with all zero eleents, which can be efficiently ipleented. he tie coplexity of D- MVH is independent of database size, it only relies on buffer size B and code length C. Extension to visual and text features In real-world applications, users ay not use all the views as query and be likely to choose parts of the. Besides the ulti-view features, we specifically consider the case of visual features and only text feature. For the query based on visual features, its hash codes can be coputed as:h v = M 1 Kt 1. Where view 1 to M 1 are the sgn αz 2 v visual features and view M is text feature. If the query only contains text feature e.g. several keywords, we can copute 1 its hash codes ash te = sgn zte M Kt. 4 Experients 4.1 Datasets We use two ulti-view iage datasets: MIR Flickr [Huiskes and Lew, 2008] and NUS-WIDE [Chua et al., 2009] to copare the hashing perforance in the online scenario. MIR Flickr contains iages collected fro Flickr. All iages are annotated with 38 class labels which are used as the ground truth. In the retrieval process, iages which share at least one sae label are considered as relevant. We select 1% iages as queries, and the rest iages are added to the database sequentially. We use 3 visual features [Guillauin et al., 2010], including 100- D Hue histogra, 1000-D SIF BoVW, 4096-D RGB histogra and 512-D GIS. Each iage is associated with text tags, thus we use 457-D BoW as text feature. NUS-WIDE contains iages collected fro Flickr, each iages is labeled by 81 concepts which can be used for evaluation. We select 1% iages as queries and the rest iages fors the streaing database. We use 3 visual features, including 500-D SIF BoVW, 64- D color histogra, 144-D color correlogra, 73-D edge direction histogra, 128-D wavelet texture and 225-D block-wise color. Each iage is associated with text tags, thus we also use 1000-D BoW as text feature. 4.2 Experiental Settings In the ipleentation of, we use Gaussian kernel for all visual features, and histogra intersection kernel for text feature. does not contain any paraeters to set. he regularization λ is set to 10 3, it is used to avoid the atrix singularity and has little influence on the results. he axiu buffer size is set to 1000 on MIR Flickr and 5000 on NUS-WIDE respectively, the threshold ρ is set to. Since our ethods does not use any supervised inforation, we copare our ethod with several representative unsupervised hashing ethods, including Online Sketching Hashing OSH [Leng et al., 2015], Collective Matrix Factorization Hashing [Ding et al., 2014], Multiview Alignent Hashing MAH [Liu et al., 2015] and Inter-edia Hashing [Song et al., 2013]. OSH cannot support any single-view or partial-view query such as text query and iage query. can only cope with two views, thus we concatenate all visual features into a visual feature. MAH cannot support visual or text query, it only supports ultiview query. supports visual and text query, but cannot support ulti-view query. Mean average precision [Song et al., 2013] is used to easure the effect of retrieval, and scores are coputed on the top 50 retrieved docuents of each query. Moreover, we evaluate the learning tie of all ethods to easure 3137

6 Proceedings of the wenty-sixth International Joint Conference on Artificial Intelligence IJCAI-17 able 1: Learning tie and code length at different sizes on NUS-WIDE. Database size OSH tie s tie s code length the efficiency of retrieval. Learning tie is coputed as the total tie of learning hash codes and functions, thus it accuulates the tie of all learning processes fro first step to current step. All the experients are conducted on a coputer with Intel CoreM i5 2.6GHz 2 processors and 12.0GB RAM Multi view Query Iage Query ext Query Seconds Experiental Results In the experients, we consider three types of queries: visual query, text query and ulti-view query. he visual query contains all visual features, text query contains only one text feature, and ulti-view query consists of all features. On MIR Flickr, iages are added to database sequentially, thus we have to evaluate the hashing perforance at different database size. More specifically, we evaluate the scores at t = {2.5, 5, 7.5, 10, 12.5, 15, 17.5, 20, 22.5, 25} he code length of DVMH is changed with the increase of database size, therefore is not appropriate to evaluate the perforance of other ethods at various code lengths. According to previous results [Ding et al., 2014], we choose the best code length as 64 bits. In order to ake a fair coparison, will not increase the code length when it reaches 64 bits. Figure 2 shows the scores of MIR Flickr at different database sizes. In iage query and text query, we copare with and. In ulti-view query, we copare with, MAH and OSH. Fro this figure we can find that obtains higher scores than other ethods. In iage and ulti-view query, the scores are increased with the data size. his phenoenon illustrates the effectiveness of our dynaic schee where code length is adaptively augented. With the augentation of code length, the perforance of can be consistently iproved. Siilar to MIR Flickr, on NUS-WIDE we evaluate the scores at different sizes, the best code length of copared ethods is set to 128 bits. he code length of will not increase when it reaches 128 bits. Figure 3 shows the scores at different database sizes. Siilar to the results of MIR Flickr, outperfors other ulti-view hashing ethods. On all types of queries, the socres of are iproved with the increase of data size. he increase of scores on NUS-WIDE are ore significant than MIR Flickr, the reason is that NUS-WIDE contains uch ore iages, thus the increase of code length are ore benefit to NUS-WIDE. Note that on NUS-WIDE, we use all 81 labels for evaluation, other works such as [Ding et al., 2014] use a reduced version which only contains 10 labels. herefore the overall scores reported in this paper are relatively lower Buffer Buffer a Scores b raining tie Figure 4: Effects of buffer size on NUS-WIDE. 4.4 Efficiency Analysis we also copare the learning tie of to online hashing ethod OSH. able 1 shows the learning tie of two ethods, and the code length of at different sizes. We can find that consues less learning tie than OSH. uses the buffer schee to avoid the frequent updating process. It is ore efficient than OSH while guarantees the hashing perforance. In addition, we can find that the code length of is dynaically changed with the increase of data size. 4.5 Effects of Buffer At last we evaluate the effects of buffer size. able 4 shows the coparison of and training tie with buffer size {500,1000,3000, 5000, 7000, 10000}. All the results are e- valuated at the final step of NUS-WIDE, where all iages are included in database. We can find that the results are relatively steady at different buffer size, which deonstrates the robustness of. Moreover, both s- cores and training tie decrease with the buffer size, thus we can choose a proper size by considering the tradeoff between scores and training tie. 5 Conclusions In this paper we propose Dynaic Multi-view Hashing D- MVH for online retrieval of streaing iage data. can adaptively augent code length to preserve ore inforation of new iage. It constructs hash codes by a dynaic dictionary, when new data cannot be effectively represented by current hash codes, can augent the dictionary to better represent this data. o avoid the frequent updating of dictionary. A buffer schee is designed and only data which are significantly different to current dictionary eleents are considered for updating. Experiental results on MIR Flickr and NUS-WIDE deonstrate the efficiency and effectiveness of copared to representative online hashing and ulti-view hashing ethods

7 Proceedings of the wenty-sixth International Joint Conference on Artificial Intelligence IJCAI-17 References [Caicedo et al., 2012] Juan C Caicedo, Jaafar BenAbdallah, Fabio A González, and Olfa Nasraoui. Multiodal representation, indexing, autoated annotation and retrieval of iage collections via non-negative atrix factorization. Neurocoputing, 761:50 60, [Cakir and Sclaroff, 2015a] Fatih Cakir and Stan Sclaroff. Adaptive hashing for fast siilarity search. In CVPR, pages , [Cakir and Sclaroff, 2015b] Fatih Cakir and Stan Sclaroff. Online supervised hashing. In ICIP, pages , [Chua et al., 2009] at-seng Chua, Jinhui ang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a realworld web iage database fro national university of singapore. In ACM CIVR, page 48, [Ding et al., 2014] Guiguang Ding, Yuchen Guo, and Jile Zhou. Collective atrix factorization hashing for ultiodal data. In CVPR, pages , [Guillauin et al., 2010] Matthieu Guillauin, Jakob Verbeek, and Cordelia Schid. Multiodal sei-supervised learning for iage classification. In CVPR, pages , [Guo et al., 2017a] Y. Guo, G. Ding, J. Han, and Y. Gao. Zero-shot learning with transferred saples. IEEE rans. Iage Processing, 2617: , [Guo et al., 2017b] Yuchen Guo, Guiguang Ding, Li Liu, Jungong Han, and Ling Shao. Learning to hash with optiized anchor ebedding for scalable retrieval. IEEE rans. Iage Processing, 263: , [Huang et al., 2013] Long-Kai Huang, Qiang Yang, and Wei-Shi Zheng. Online hashing. In IJCAI, [Huiskes and Lew, 2008] Mark J Huiskes and Michael S Lew. he ir flickr retrieval evaluation. In ACM ICMR, pages 39 43, [Ki and Choi, 2013] Saehoon Ki and Seungjin Choi. Multiview anchor graph hashing. In Acoustics, Speech and Signal Processing ICASSP, 2013 IEEE International Conference on, pages , [Kuar and Udupa, 2011] Shaishav Kuar and Raghavendra Udupa. Learning hash functions for cross-view siilarity search. In IJCAI, page 1360, [Leng et al., 2015] Cong Leng, Jiaxiang Wu, Jian Cheng, Xiao Bai, and Hanqing Lu. Online sketching hashing. In CVPR, pages , [Lin et al., 2017a] Z. Lin, G. Ding, J. Han, and L. Shao. End-to-end feature-aware label space encoding for ultilabel classification with any classes. IEEE rans. Neural Networks and Learning Systes, [Lin et al., 2017b] Z. Lin, G. Ding, J. Han, and J. Wang. Crossview retrieval via probability-based seantics-preserving hashing. IEEE rans. Cybernetics, [Liu et al., 2015] Li Liu, Mengyang Yu, and Ling Shao. Multiview alignent hashing for efficient iage search. Iage Processing, IEEE ransactions on, 243: , [Liu et al., 2017] Li Liu, Zijia Lin, Ling Shao, Fuin Shen, Guiguang Ding, and Jungong Han. Sequential discrete hashing for scalable cross-odality siilarity retrieval. IEEE rans. Iage Processing, 261: , [Mu et al., 2010] Yadong Mu, Jialie Shen, and Shuicheng Yan. Weakly-supervised hashing in kernel space. In CVPR, pages , [Rasiwasia et al., 2010] Nikhil Rasiwasia, Jose Costa Pereira, E- anuele Coviello, Gabriel Doyle, Gert RG Lanckriet, Roger Levy, and Nuno Vasconcelos. A new approach to cross-odal ultiedia retrieval. In ACM MM, pages , [Song et al., 2013] Jingkuan Song, Yang Yang, Yi Yang, Zi Huang, and Heng ao Shen. Inter-edia hashing for large-scale retrieval fro heterogeneous data sources. In ACM SIGMOD, pages , [Wu et al., 2014] Fei Wu, Zhou Yu, Yi Yang, Siliang ang, Yin Zhang, and Yueting Zhuang. Sparse ulti-odal hashing. Multiedia, IEEE ransactions on, 162: , [Wu et al., 2015] Botong Wu, Qiang Yang, Wei-Shi Zheng, Yizhou Wang, and Jingdong Wang. Quantized correlation hashing for fast cross-odal search. In IJCAI, [Wu et al., 2017] Gengshen Wu, Li Liu, Yuchen Guo, Guiguang D- ing, Jungong Han, Jialie Shen, and Ling Shao. Unsupervised deep video hashing with balanced rotation. In IJCAI, [Xie et al., 2016] Liang Xie, Jialie Shen, and Lei Zhu. Online cross-odal hashing for web iage retrieval. In AAAI, pages , [Yu et al., 2014] Zhou Yu, Fei Wu, Yi Yang, Qi ian, Jiebo Luo, and Yueting Zhuang. Discriinative coupled dictionary hashing for fast cross-edia retrieval. In ACM SIGIR, pages , [Zhang and Li, 2014] Dongqing Zhang and Wu-Jun Li. Large-scale supervised ultiodal hashing with seantic correlation axiization. In AAAI, pages , [Zhang et al., ] Yin Zhang, Weiing Lu, Yang Liu, and Fei Wu. Kernelized sparse hashing for scalable iage retrieval. Neurocoputing, 172. [Zhang et al., 2011] Dan Zhang, Fei Wang, and Luo Si. Coposite hashing with ultiple inforation sources. In ACM SIGIR, pages , [Zhen and Yeung, 2012a] Yi Zhen and Dit-Yan Yeung. Coregularized hashing for ultiodal data. In NIPS 25, pages [Zhen and Yeung, 2012b] Yi Zhen and Dit-Yan Yeung. A probabilistic odel for ultiodal hash function learning. In ACM SIGKDD, pages , [Zhou et al., 2014] Jile Zhou, Guiguang Ding, and Yuchen Guo. Latent seantic sparse hashing for cross-odal siilarity search. In ACM SIGIR, pages , [Zhu et al., 2017] Lei Zhu, Jialie Shen, Liang Xie, and Zhiyong Cheng. Unsupervised visual hashing with seantic assistant for content-based iage retrieval. IEEE rans. Knowl. Data Eng., 292: ,