6004 IEEE TRANSACTIONS ON IMAGE PROCESSING,VOL.27.NO.12,DECEMBER 2018 0 -DH 02 00 Recall (a) 6 (c) (d) 09 07 045 4 1 09 DHN +D 0 。=D50E 03 0 DPSI 一IN DHN 0 Recall 6 02 Recall (e) (f) (g) (h) 08 04 30 0 DHN DG用 =DD 。=sDH 03 DPSH =DP5 H 0 -DHN 02 08 02 02 0. (i) ) (k) () 08 03 -00SII +一山5川 DPSH 0.3 年0H --DHN 0 ◆DDB =DGD用 0. 05 0 02 02 04 06 0 (m) (n) (o) (p) Fig.2.Performance of precision-recall curve on four datasets.The four sub-figures in each row are the precision-recall curves for 12 bits,24 bits,32 bits and 48 bits,respectively.(a)12 bits @CIFAR-10.(b)12 bits @SVHN.(c)12 bits @NUS-WIDE.(d)12 bits @ClothingIM.(e)24 bits @CIFAR-10. (f)24 bits @SVHN.(g)24 bits @NUS-WIDE.(h)24 bits @ClothingIM.(i)32 bits @CIFAR-10.(j)32 bits @SVHN.(k)32 bits @NUS-WIDE.(1)48 bits @Clothing1M.(m)48 bits @CIFAR-10.(n)48 bits @SVHN.(o)48 bits @NUS-WIDE.(p)48 bits @ClothingIM. By comparing NDH,COSDISH,SDH and FastH to LFH. to directly guide deep feature learning procedure but other we can find that discrete supervised hashing can outper- discrete supervised hashing methods do not have deep feature form relaxation-based continuous hashing,which means that learning ability.The main difference between our DDSH and discrete coding procedure is able to learn more optimal other deep hashing methods is that DDSH adopts supervised binary codes.By comparing feature learning based deep information to directly guide the discrete coding procedure but hashing methods,i.e.,DDSH,DSDH,DPSH,DHN and DSH, other deep hashing methods do not have this property.Hence. to non-deep hashing methods,we can find that feature learning the experimental results successfully demonstrate the motiva- based deep hashing can outperform non-deep hashing because tion of DDSH,i.e.,utilizing supervised information to directly deep supervised hashing can perform deep feature learning guide both deep feature learning procedure and discrete coding compared with non-deep hashing methods.This experimental procedure can further improve retrieval performance in real result demonstrates that deep supervised hashing is a more applications. compatible architecture for hashing learning. Furthermore,we select four best baselines,i.e.,DSDH, The main difference between our proposed DDSH and other DPSH,DSH and DHN,to compare the precision-recall and discrete supervised hashing methods like COSDISH,SDH top-k precision results.We report the precision-recall curve and FastH is that our DDSH adopts supervised information on all four datasets in Figure 2.We can see that the6004 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 27, NO. 12, DECEMBER 2018 Fig. 2. Performance of precision-recall curve on four datasets. The four sub-figures in each row are the precision-recall curves for 12 bits, 24 bits, 32 bits and 48 bits, respectively. (a) 12 bits @CIFAR-10. (b) 12 bits @SVHN. (c) 12 bits @NUS-WIDE. (d) 12 bits @Clothing1M. (e) 24 bits @CIFAR-10. (f) 24 bits @SVHN. (g) 24 bits @NUS-WIDE. (h) 24 bits @Clothing1M. (i) 32 bits @CIFAR-10. (j) 32 bits @SVHN. (k) 32 bits @NUS-WIDE. (l) 48 bits @Clothing1M. (m) 48 bits @CIFAR-10. (n) 48 bits @SVHN. (o) 48 bits @NUS-WIDE. (p) 48 bits @Clothing1M. By comparing NDH, COSDISH, SDH and FastH to LFH, we can find that discrete supervised hashing can outper￾form relaxation-based continuous hashing, which means that discrete coding procedure is able to learn more optimal binary codes. By comparing feature learning based deep hashing methods, i.e., DDSH, DSDH, DPSH, DHN and DSH, to non-deep hashing methods, we can find that feature learning based deep hashing can outperform non-deep hashing because deep supervised hashing can perform deep feature learning compared with non-deep hashing methods. This experimental result demonstrates that deep supervised hashing is a more compatible architecture for hashing learning. The main difference between our proposed DDSH and other discrete supervised hashing methods like COSDISH, SDH and FastH is that our DDSH adopts supervised information to directly guide deep feature learning procedure but other discrete supervised hashing methods do not have deep feature learning ability. The main difference between our DDSH and other deep hashing methods is that DDSH adopts supervised information to directly guide the discrete coding procedure but other deep hashing methods do not have this property. Hence, the experimental results successfully demonstrate the motiva￾tion of DDSH, i.e., utilizing supervised information to directly guide both deep feature learning procedure and discrete coding procedure can further improve retrieval performance in real applications. Furthermore, we select four best baselines, i.e., DSDH, DPSH, DSH and DHN, to compare the precision-recall and top-k precision results. We report the precision-recall curve on all four datasets in Figure 2. We can see that the