IEEE Computational Intelligence Magazine - May 2020 - 73

significantly better or significantly
worse than the compared method. The
symbol "="indicates the performance of
FLGP is similar to the compared method. In Tables III and IV, each small
block lists the results on one data set,
and the maximum classification accuracy is highlighted in bold. The final row

of each block summarizes the overall
results of the significance test.
A. Overall Classification Performance

As mentioned in Section IV-B, the GPGLF and the other four GP-based
benchmark methods are only used for
comparisons on binary classification

tasks. Thus, there are 16 benchmark
methods on FEI_1, FEI_2 and VGDB
and 11 benchmark methods on the
remaining data sets. From the final rows
of Tables III and IV, it is obvious that
FLGP achieves significantly better or
similar performance in almost all the
comparisons. Specifically, FLGP gains 97

TABLE III Classification accuracy (%) of the proposed FLGP approach and sixteen benchmark methods on three binary data sets:
FEI_1, FEI_2 and VGDB.
FEI_1

FEI_2

VGDB

METHODS

MAX

MEAN±STD.

MAX

MEAN±STD.

MAX

MEAN ± STD.

2TGP

96.0

88.1 ± 6.2 +

94.0

85.5 ± 6.0 +

63.9

61.6 ± 1.5 +

DIF + GP

80.0

56.7 ± 6.9 +

72.0

60.3 ± 8.4 +

68.7

61.4 ± 3.5 +

Hist + GP

70.0

48.9 ± 7.2 +

60.0

48.8 ± 6.1 +

84.3

76.0 ± 2.6 =

uLBP+GP

66.0

50.9 ± 7.5 +

72.0

48.7 ± 7.9 +

79.5

69.4 ± 4.4 +

GP-GLF

96.0

89.1 ± 3.9 +

92.0

82.5 ± 5.8 +

74.6

65.1 ± 4.7 +

DIF

74.0

61.1 ± 4.9 +

72.0

62.8 ± 6.1 +

66.3

55.6 ± 10.3 +

Hist

54.0

48.1 ± 3.4 +

54.0

50.1 ± 2.5 +

62.7

62.2 ± 0.8 +

GLCM

50.0

49.7 ± 0.7 +

54.0

50.1 ± 0.7 +

62.7

53.3 ± 9.8 +

Gabor

82.0

71.6 ± 7.9 +

74.0

65.7 ± 5.1 +

63.9

56.0 ± 8.3 +

SIFT

82.0

82.0 ± 0.0 +

78.0

78.0 ± 0.0 +

60.2

60.2 ± 0.0 +

HOG

94.0

94.0 ± 0.0 +

88.0

88.0 ± 0.0 +

57.8

57.2 ± 0.7 +

LBP

68.0

62.5 ± 3.5 +

66.0

57.6 ± 3.6 +

84.3

80.6 ± 3.2 -

uLBP

64.0

56.9 ± 5.2 +

56.0

51.9 ± 2.3 +

81.9

71.5 ± 8.1 =

LeNet-5

98.0

94.4 ± 2.0 =

94.0

90.8 ± 1.8 +

65.1

58.1 ± 4.8 +

CNN-5

98.0

95.6 ± 1.5 =

90.0

85.0 ± 3.0 +

65.1

61.5 ± 2.1 +

CNN-8

98.0

94.2 ± 2.1 =

94.0

90.0 ± 2.3 +

61.5

56.9 ± 4.6 +

FLGP

98.0

95.8 ± 3.2

100

93.3 ± 3.8

81.9

74.9 ± 3.7

OVERALL

13 +, 3 =

16+

13 +, 2 =, 1 -

TABLE IV Classification accuracy (%) of the proposed FLGP approach and eleven benchmark methods on five multi-class data sets:
ORL, JAFFE, KTH, EYALE, and SCENE.
ORL
METHODS

MAX

MEAN ±
STD.

JAFFE
MAX

MEAN ±
STD.

KTH
MAX

EYALE

MEAN ±
STD.

MAX

MEAN ±
STD.

SCENE
MAX

MEAN ±
STD.

DIF

85.0

85.0 ± 0.0 +

35.6

35.6 ± 0.0 +

56.2

56.2 ± 0.0 +

26.4

26.4 ± 0.0 +

33.5

33.5 ± 0.0 +

Hist

97.5

97.5 ± 0.0 +

19.2

19.2 ± 0.0 +

51.4

51.4 ± 0.0 +

11.0

11.0 ± 0.0 +

21.2

21.2 ± 0.0 +

GLCM

2.5

2.5 ± 0.0 +

15.1

15.1 ± 0.0 +

23.3

23.3 ± 0.0 +

5.1

5.0 ± 0.1 +

13.9

13.9 ± 0.0 +

Gabor

59.2

57.1 ± 0.9 +

46.6

43.2 ± 1.6 +

44.3

42.9 ± 0.7 +

36.7

36.3 ± 0.2 +

22.7

22.3 ± 0.2 +

SIFT

98.3

98.3 ± 0.0 +

74.0

74.0 ± 0.0 +

81.4

81.4 ± 0.0 +

88.4

88.4 ± 0.0 +

63.1

63.1 ± 0.0 +

HOG

96.7

96.7 ± 0.0 +

72.6

72.6 ± 0.0 +

51.4

51.4 ± 0.2 +

74.3

74.3 ± 0.0 +

30.8

30.8 ± 0.0 +

LBP

87.5

87.5 ± 0.0 +

21.9

21.9 ± 0.0 +

87.6

87.6 ± 0.0 +

46.4

46.4 ± 0.0 +

62.5

62.5 ± 0.0 +

uLBP

94.2

94.2 ± 0.0 +

23.3

23.3 ± 0.0 +

81.0

81.0 ± 0.0 +

56.1

56.1 ± 0.0 +

66.1

66.1 ± 0.0 +

LeNet

93.3

89.9 ± 1.9 +

79.5

68.9 ± 7.0 +

78.6

72.0 ± 6.4 +

92.4

89.4 ± 1.6 +

54.2

51.1 ± 2.4 +

CNN-5

97.5

96.3 ± 0.8 +

80.8

78.9 ± 1.3 +

84.3

81.5 ± 1.8 +

99.3

98.6 ± 0.5 +

58.9

55.4 ± 1.4 +

CNN-8

96.7

94.2 ± 1.8 +

61.6

52.5 ± 6.2 +

82.4

80.7 ± 1.5 +

90.9

88.2 ± 1.0 +

69.2

66.2 ± 2.0 +

FLGP

100

99.6 ± 0.7

91.8

81.1 ± 4.7

95.7

94.5 ± 0.9

99.8

99.2 ± 0.4

77.0

75.2 ± 0.7

OVERALL

11 +

11 +

11 +

11 +

11 +

MAY 2020 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE

73
IEEE Computational Intelligence Magazine - May 2020

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - May 2020
