Table 3 The retrieval performance of the top 500 ranked images in MAP (%) for difference approaches on NUS-WIDE and Flickr

Datasets

NUS-WIDE

Methods

SR

SR+BOW

SR+GlV

SR+lc

SR+tm

SR+gp

SR+ctg

SR+id

SR+all

PAMIR-PR (shallow)

27.6

28.3

30.6

27.9

26.9

28.0

28.2

27.9

31.4

PAMIR-PR (deep)

29.1

30.2

31.3

28.9

28.2

29.3

29.4

29.4

32.6

MMNN-PR (shallow)

26.3

26.9

27.8

26.3

25.5

26.9

27.3

25.9

29.5

MMNN-PR (deep)

28.2

29.0

30.4

29.4

28.3

29.1

30.1

28.1

31.3

SCM-PR (shallow)

21.5

21.9

23.7

21.5

20.1

21.9

22.1

21.2

24.4

SCM-PR (deep)

23.4

23.8

25.6

23.9

23.2

23.7

24.2

23.0

25.5

CMOS _lg-PR

36.4

36.8

38.4

37.0

36.7

37.1

37.6

36.0

38.8

Deep-PR

38.7

38.9

40.9

39.8

39.3

39.8

39.9

38.5

41.6

OSIR _s (shallow)

31.7

32.2

32.9

32.0

31.8

32.1

32.5

31.9

32.6

OSIR _s (deep)

35.9

36.7

38.3

36.4

36.1

36.4

36.9

36.3

38.9

OSIR (shallow)

38.5

39.4

40.3

39.1

38.4

38.7

40.4

39.1

43.8

OSIR (deep)

40.3

41.2

41.8

41.1

40.1

40.6

41.3

40.7

44.4

Datasets

Flickr

Methods

SR

SR+BOW

SR+GlV

SR+lc

SR+tm

SR+gp

SR+ctg

SR+id

SR+all

PAMIR-PR (shallow)

12.4

13.2

13.9

12.8

12.0

13.1

14.0

12.3

14.5

PAMIR-PR (deep)

15.1

15.9

16.7

15.5

15.0

16.1

17.2

15.2

17.9

MMNN-PR (shallow)

12.1

12.9

13.4

12.6

11.8

13.2

13.6

12.0

14.0

MMNN-PR (deep)

14.8

15.3

15.9

15.5

14.2

16.1

16.4

14.9

16.9

SCM-PR (shallow)

5.8

6.2

6.7

5.9

5.5

6.5

6.9

6.3

7.6

SCM-PR (deep)

7.2

7.8

8.6

7.4

7.8

8.1

8.9

7.1

9.4

CMOS _lg-PR

19.4

20.1

21.6

19.9

19.7

20.3

21.4

19.2

23.7

Deep-PR

19.9

20.3

21.8

20.2

20.0

20.9

22.0

19.5

24.1

OSIR _s (shallow)

14.3

15.1

15.8

14.9

14.4

15.8

16.2

15.1

18.1

OSIR _s (deep)

16.5

17.1

18.3

16.9

16.7

17.2

17.8

17.1

20.3

OSIR (shallow)

18.1

19.3

19.8

19.2

18.3

20.2

20.3

18.9

23.4

OSIR (deep)

20.6

21.2

21.9

21.4

20.7

21.7

22.7

21.4

25.3