Article
Nonlinear Fusion of Multispectral Citrus Fruit Image
Data with Information Contents
Peilin Li
1
, Sang-Heon Lee
1
, Hung-Yao Hsu
1
and Jae-Sam Park
2,
*
1
School of Engineering, University of South Australia, Mawson Lakes 5095, Australia;
liypl001@mymail.unisa.edu.au (P.L.); Sang-Heon.Lee@unisa.edu.au (S.-H.L.);
Hung-Yao.Hsu@unisa.edu.au (H.-Y.H.)
2
Department of Electronics Engineering, Incheon National University, 119 Academy Road, Yeon Su Gu,
Incheon 22012, Korea
* Correspondence: jaepark@inu.ac.kr; Tel.: +82-32-835-8457
Academic Editor: Gonzalo Pajares Martinsanz
Received: 6 November 2016; Accepted: 9 January 2017; Published: 13 January 2017
Abstract:
The main issue of vison-based automatic harvesting manipulators is the difficulty in the
correct fruit identification in the images under natural lighting conditions. Mostly, the solution has
been based on a linear combination of color components in the multispectral images. However,
the results have not reached a satisfactory level. To overcome this issue, this paper proposes a
robust nonlinear fusion method to augment the original color image with the synchronized near
infrared image. The two images are fused with Daubechies wavelet transform (DWT) in a multiscale
decomposition approach. With DWT, the background noises are reduced and the necessary image
features are enhanced by fusing the color contrast of the color components and the homogeneity of
the near infrared (NIR) component. The resulting fused color image is classified with a C-means
algorithm for reconstruction. The performance of the proposed approach is evaluated with the
statistical F measure in comparison to some existing methods using linear combinations of color
components. The results show that the fusion of information in different spectral components has the
advantage of enhancing the image quality, therefore improving the classification accuracy in citrus
fruit identification in natural lighting conditions.
Keywords: image fusion; entropy filter; multiscale decomposition; wavelet transform; clustering
1. Introduction
For the robotic harvesting manipulator used in the horticultural industry, the main technique
used to identify the location of fruits is a vision system. Since a vision system was proposed in this
field [
1
], various sensor schemes have been practiced utilizing the intensity, the spectral information,
or the laser range finder [
2
]. The successful fruit identification rate has been reported between 70% and
90% with a variation in laboratory conditions. Some issues are still to be solved before the widespread
commercial use of the automatic harvesting manipulator. Normally, a color intensity thresholding
with a certain filter has been used to contrast the salient features of an image [
3
]. In practice, the
image data acquired by a single sensor is degraded since the imaging sensors have certain physical
limitations. On the other hand, the light spectrum is potentially affected by multiple factors in an open
unstructured environment [
4
]. In order to study the features from different wavebands on spectral
coordinates, the hyperspectral techniques have been used to capture the necessary information from a
wide range of the light spectrum. The study of the statistics of image segments obtained from different
wavebands, in particular, techniques involving fuzzy wavebands, may provide the feature references
necessary for the development of a machine vision system [
5
]. However, the distribution of the spectra
on the segments of some components from certain spectral coordinates can be fuzzy and hence make
Sensors 2017, 17, 142; doi:10.3390/s17010142 www.mdpi.com/journal/sensors
