IEEE Computational Intelligence Magazine - May 2020 - 76

using Region_R from the input image.
From each detected region, it extracts
128 SIFT features using L_SIFT .
Together with the extracted 20 DIF features from the whole image by G_DIF,
the solution is able to produce 276 features from an input image.
From Fig. 8, it can be seen that the
Region_S function detects the chin
and mouth area of the left face and
the Region_S function detects a similar area of the right face. It is obvious
that the two regions capture the most
discriminative information between
the two different classes. For example,
the detected area contains the teeth in
the happy face image but not in the
natural face image and captures the
difference in mouth shapes between
the two expressions.
This example solution finds a combination of local SIFT features and global
DIF features for classification. The combined features are more effective for
classification than the individual global
DIF and SIFT features. The traditional
method using DIF features only
achieves a maximum accuracy of 72%
and the method using SIFT features
only achieves a maximum accuracy of
78% on FEI_2. The example solution
improves the classification performance
by detecting the regions of interest and
extracting meaningful local features
from the detected regions. The analysis
shows that FLGP detects informative
regions and extracts discriminative local
and global features for classification.

2) Analysis on the Feature
Extraction Functions
To further analyze FLGP, ten best
FLGP solutions of each run (totally
300 solutions) on each data set are
recorded. The frequency of the ten feature extraction functions in the global
(G_DIF, G_Hist, G_SIFT, G_HOG and
G_uLBP) and local (L_DIF, L_Hist,
L_SIFT, L_HOG, and L_uLBP ) scenarios in these solutions are ranked. The
results of the ranking are listed in Table V,
where 1 indicates the most frequently
used function and 10 indicates the least
frequently used function in these solutions. From Table V, it is clear that the
frequency-rank of the feature extraction
functions varies with the data sets. Specifically, the L_SIFT function is the
most frequently used on FEI_1, FEI_2,
JAFFE, and EYALE, the G_uLBP function is the most frequently used on
VGDB, KTH and SCENE, and the
G_SIFT function is the most frequently
used on ORL. Moreover, the frequently
used functions on one data set may be
less frequently used on the other data
sets. For example, L_SIFT is the most
frequently used on four data sets but it is
the least frequently used on ORL, KTH
and SCENE. The G_Hist function is
the least frequently used on FEI_1,
FEI_2 and JAFFE, but it is frequently
used on VGDB. This confirms the difficulty of feature extraction as different data sets need various types of
features. In contrast, FLGP automatically finds the best feature extraction

TABLE V Ranking of all the feature extraction functions in 300 best-of-the-run
programs of FLGP on each data set.

76

FUNCTION

FEI_1

FEI_2

VGDB

ORL

JAFFE

KTH

EYALE

SCENE

G_DIF

5

9

7

7

7

5

8

4

G_Hist

10

10

3

5

9

6

5

6

G_SIFT

8

5

8

1

2

2

3

2

G_HOG

3

4

10

4

5

7

9

9

G_uLBP

9

7

1

2

10

1

4

1

L_DIF

6

6

5

9

3

4

7

7

L_Hist

4

8

2

6

8

3

6

8

L_SIFT

1

1

9

10

1

10

1

10

L_HOG

7

3

4

8

4

9

10

5

L_uLBP

2

2

6

3

6

8

2

3

IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | MAY 2020

methods or combinations of them to
extract features.
It can be seen from Table V that
FLGP learns more local features than
global features on face data sets, i.e.,
FEI_1, FEI_2, JAFFE, and EYALE,
which confirms that local features are
more effective for object classification. In
contrast, FLGP learns more global features than local features on the nonobject data sets, i.e., VGDB, KTH and
SCENE, as global features are more
effective. The analysis shows that FLGP
learns the best feature extraction functions or combinations of them to extract
effective global and/or local features for
image classification.
VII. Conclusions

The aim of this paper was to develop an
effective GP-based approach that uses
existing well-developed feature descriptors
to learn global and/or local features for
different image classification tasks. This
goal was successfully achieved by proposing the FLGP approach with a new program structure, a new function set and a
new terminal set and examining FLGP on
eight different image data sets of varying
difficulty. Specifically, five existing feature
descriptors, i.e., Hist, DIF, SIFT, HOG, and
uLBP, were employed in FLGP as functions in the global and local scenarios.
FLGP was compared with a number of
benchmark methods to show its effectiveness. Example solutions of FLGP were
visualized and analyzed to show the good
interpretability of the solutions.
Experimental results demonstrated
that FLGP achieved significantly better
performance in almost all the comparisons on different image classification tasks.
The results confirmed the capability of
FLGP on automatically learning effective
global and/or local features for achieving
high classification accuracy. The comparisons of FLGP and GP-GLF showed that
FLGP significantly improved the classification performance by using a flexible
program structure, a new function set and
a new feature learning process. The comparisons of FLGP and the other four GPbased methods suggested that it is more
effective to automatically learn a set of
high-level features from raw pixels than
IEEE Computational Intelligence Magazine - May 2020

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