Feature Selection From Hyperspectral Imaging Through Discriminant Function For Guava Defects Detection
Loading...
Date
2017-06
Authors
Tan, Suo Ching
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Sains Malaysia
Abstract
Development of technology makes hyperspectral imaging commonly used for
defects detection. In this research, a hyperspectral imaging system was setup with the
study of lighting and compare the selected feature through different feature selection
methods. Light Emitting Diode (LED) and fluorescent light sources was examined and
found that both light sources having an almost identical reflectance pattern. The peak
intensity of the fluorescent light source is 4.5 times higher than this particular LED.
By considering the signal to noise ratio, fluorescent light source was selected for
subsequent experiments. Sequential feature selection with linear discriminant (LD)
and quadratic discriminant (QD) function were used to select several features for
further processing. Different training methods were used in this research. The first
method which refers to conventional method (CM) is to feed the training sample as a
whole dataset into the selection. The second method is to feed the training sample by
separate the image into two parts (TPTM) which refer to a higher reflectivity area and
lower reflectivity area. Four matrixes of validation were done, which are SFS-LD-CM,
SFS-LD-TPTM, SFS-QD-CM and SFS-QD-TPTM. These methods were evaluated by
F1-score. With 48 defect areas, experiment shown that F1-score of SFS-LD-TPTM
has a higher detection rate which is 0.8 in F1-score, compared with 0.78 by SFS-QDCM
and 0.72 for both SFS-LD-CM and SFS-QD-TPTM. This concludes that by
selecting the features with two training samples able to minimize the impact of uneven
illumination for this particular system and hence having higher correct detection rate
in guava defects detection.
Description
Keywords
Feature selection from Hyperspectral imaging through , Discriminant function For guava defects detection