Recent Advances and Applications of Hyperspectral Imaging for Fruit and Vegetable Quality Assessment

TL;DR: The different technologies available to acquire the images and their use for the non-destructive inspection of the internal and external features of these products are explained, with details of the statistical techniques most commonly used for this task.

Abstract: Hyperspectral imaging systems are starting to be used as a scientific tool for food quality assessment. A typical hyperspectral image is composed of a set of a relatively wide range of monochromatic images corresponding to continuous wavelengths that normally contain redundant information or may exhibit a high degree of correlation. In addition, computation of the classifiers used to deal with the data obtained from the images can become excessively complex and time-consuming for such high-dimensional datasets, and this makes it difficult to incorporate such systems into an industry that demands standard protocols or high-speed processes. Therefore, recent works have focused on the development of new systems based on this technology that are capable of analysing quality features that cannot be inspected using visible imaging. Many of those studies have also centred on finding new statistical techniques to reduce the hyperspectral images to multispectral ones, which are easier to implement in automatic, non-destructive systems. This article reviews recent works that use hyperspectral imaging for the inspection of fruit and vegetables. It explains the different technologies available to acquire the images and their use for the non-destructive inspection of the internal and external features of these products. Particular attention is paid to the works aimed at reducing the dimensionality of the images, with details of the statistical techniques most commonly used for this task.

Summary

1. Introduction

• Most of these factors have traditionally been assessed by visual inspection or destructive sampling performed by trained operators, but currently many of them, particularly the external ones, can be estimated with commercial vision systems (Cubero et al., 2011).
• A multispectral vision system developed by Kleynen et al. (2005) included four wavelength bands in the visible/NIR range for sorting apples cv. Jonagold based on the presence of defects.
• This work was later enhanced by extracting several features from defective skins with the aim of classifying the fruit in different categories (Unay et al., 2011).
• Beyond multispectral imaging, the use of hyperspectral sensors makes it possible to conduct a more sophisticated analysis of the scene by acquiring a set of images corresponding to particular wavelengths, normally in the visible and NIR part of the electromagnetic spectrum.

2. Technologies for hyperspectral image acquisition

• The essential elements for constructing hyperspectral imaging systems include light sources, wavelength selection devices, and area detectors (Sun, 2010).
• Depending on the technology used, the selection of the wavelengths can be performed by dispersing the incident radiation into its individual wavelength or blocking the radiation in such a way that only the desired wavelength reaches the detector.
• The most frequently used are usually imaging spectrographs, liquid crystal tunable filters (LCTF) and, to a lesser extent, acousto-optic tunable filters (AOTF).
• There are also other kinds of equipment that have been developed for the acquisition of reflectance hyperspectral images (Kim et al., 2001).

2.1. Liquid crystal tunable filters

• An LCTF is a solid-state instrument that uses electronically controlled liquid crystal cells to transmit light with a selectable wavelength, while excluding all others.
• The LCTF is based on Lyot filters, which are built from a series of optical stages, each consisting of a combination of a birefringent retarder (an optical property of a material that causes the polarisations of light to travel at different speeds) and a liquid crystal layer sandwiched between two parallel polarisers.
• Typically, it takes tens of milliseconds to switch from one wavelength to another, which is far longer than the response time of the AOTF.
• The main characteristics included two liquid crystal filters, with spectral ranges of 400 nm to 720 nm, and 650 nm to 1100 nm respectively.
• Spectral images from Red Delicious and Golden Delicious apples were acquired from 650 to 1000 nm in increments of 10 nm. Gómez-Sanchis et al. (2008a) studied the feasibility of an LCTF hyperspectral system for detecting decay in citrus fruits in the early stages of infection using halogen lighting instead of the traditional inspection using UV lighting.

2.2. Acousto-optic tunable filters

• In recent years, technology based on AOTF has grown, thereby providing an alternative to LCTF and to imaging spectrographs (Vila et al., 2005), and its use is starting to be introduced for optimising agricultural and chemical processes (Bei et al., 2004).
• Jiménez et al. (2008) used an AOTF to obtain the spectrum of olive oil from inside a horizontal centrifugal decanter.
• The acoustic waves change the refractive index of the crystal by compressing and relaxing the crystal lattice.
• Therefore, the wavelength of the diffracted beam is controlled by changing the frequency of the RF source (Vila-Francés et al., 2011).
• Since AOTF is an advanced electronically tunable filter, it includes important features similar to those to be found in LCTF, such as accessibility to random wavelengths, flexible controllability, high spectral resolution, fast wavelength switching, wide spectral range, narrow bandwidth, and a relatively large optical aperture.

2.3. Imaging spectrographs

• An imaging spectrograph is an optical device that is capable of dispersing incident broadband light into different wavelengths instantaneously on an area detector (e.g. a CCD detector).
• The light from a scanning line is dispersed into different wavelengths and they are projected onto the area detector, creating a special twodimensional image: one dimension represents spatial information and the other the spectral dimension.
• Therefore, it is not possible to acquire an entire image without properly synchronising the image acquisition with the movement of the object.
• Some examples where this technology is well described include ElMasry et al. (2008), where a hyperspectral imaging system based on a spectrograph was used in the spectral region between 400 and 1000 nm for early detection of bruises on different background colours of apples cv. McIntosh.
• Polder et al. (2003) used an imaging spectrograph (393-710 nm) to estimate lycopene and chlorophyll contents, which play a role in the ripening of tomatoes.

3. Most commonly used statistical techniques

• Having a large number of bands is of great interest but also increases the complexity of the analysis of the information.
• This technique is widely used in hyperspectral imaging, as it is considered a powerful and robust tool for obtaining an overview of such complex data and for reducing the large dimension of the data provided by the hyperspectral images.
• The study showed that both methods gave very similar results for the detection of disease, fungal contamination, bruises and soil contamination on apples.
• The PCA technique has been widely applied for data reduction.
• Later, Xing et al. (2007a) used PCA in the same spectral region to reduce the number of bands for separating stem-end/calyx regions from true bruises on apples cv. Golden Delicious and cv. Jonagold.

3.2. Partial least squares

• PLS regression is an unsupervised statistical method used when not only a data array coming from X data is available, but also a Y array that the authors want to predict from their X data.
• Moreover, PLS analysis is related to PCA.
• PLS models were developed between the average reflectance spectra and the measured quality parameters in order to predict quality parameters.
• In order to study ripening in tomatoes, Polder et al. (2004) analysed concentrations of different compounds using HPLC and by analysing spectral images using PLS regression at the pixel level and at the tomato level.
• It was found that the PLS-DA models that were developed were capable of satisfactorily identifying undamaged regions, casing soil and enzymatically damaged areas on mushrooms from the validation sets.

3.3. Linear discriminant analysis

• Discriminant analysis is a statistical technique for classifying objects into mutually exclusive groups based on a set of measurable features of the objects, which in the case of hyperspectral images are normally spectral features.
• This supervised method is focused on maximizing the ratio of the variance between groups and variance within groups (Jobson, 1992).
• The discriminant functions from LDA were able to separate the pixels and classify the objects as tuber or clod under wet and dry conditions with higher rates than with just colour information.
• The hyperspectral images analysed using LDA also offered better results than the traditional RGB systems.
• The reasonably low misclassification rates obtained for classification of undamaged mushrooms and mushrooms just after thawing highlights the high potential of hyperspectral imaging combined with PCA to reduce the original data dimensional space and LDA for the early identification of mushrooms subjected to freeze damage.

3.4. Artificial neural networks

• An ANN is a non-linear statistical data-modelling tool that attempts to mimic the fault-tolerance and capacity to learn of biological neural systems by modelling the low-level structure of the brain.
• The most popular ANN is the multilayer perceptron (MLP), which is a feedforward ANN model that maps sets of input data onto a set of appropriate outputs, and consists of multiple layers of nodes on a directed graph that is fully connected from one layer to the next.
• ANN is a commonly used pattern recognition tool in hyperspectral image processing because of the fact that it is capable of handling a large amount of heterogeneous data with considerable flexibility and due to its non-linear property (Plaza et al., 2009).
• A combination of PCA and ANN was also used by Bennedsen et al. (2007) to detect surface defects on apples cv. Golden Delicious.
• Each set consisted of two categories based on the arrangement of the images: ‘vertical’ and ‘horizontal’.

4. Dimensionality reduction and selection of spectral features

• With a spectral resolution of about 5 nm, a system working between 400 and 1000 nm could acquire about 120 images , which normally contain redundant information or may exhibit a high degree of correlation.
• Methods for reducing the dimensionality can be divided into feature selection and feature extraction.
• PLS and stepwise discriminant analysis were used to reduce data dimensionality and to select the effective wavelengths.
• The images at the selected wavelengths were averaged, thereby creating a new image that was the basis for bruise area identification using a multilevel adaptive thresholding method.
• A different approach was taken by Ariana and Lu (2010), who used hyperspectral imaging under the transmittance mode to select important wavebands that can be used in a further development of an in-line inspection system to detect internal defects in pickling cucumbers and whole pickles.

5. Estimation of fruit quality

• Hyperspectral imaging has recently emerged as a powerful inspection tool for quality assessment of fruits and vegetables.
• The quality of a piece of fruit or vegetable is defined by several attributes that determine its marketability and shelf life.
• Quality assessment is therefore one of the most important goals of the highly competitive food industry.
• Even though such techniques offer important advantages like real-time operation, lower cost or simulation of human processes, they also have some limitations, the main one being the fact that the human eye is restricted to the visible part of the electromagnetic spectrum and misses important information that is outside these limits.
• Therefore, to expand quality inspection beyond human limitations, it is necessary to employ instrumental measurements such as hyperspectral imaging (Sun, 2010).

5.1. Estimation of external quality parameters

• Detection of skin defects is one of the most widesprad uses of hyperspectral imaging in the inspection of fruits and vegetables, since the perceived quality is highly associated with a good appearance of the product.
• The contour plots for the first PC score images were used to distinguish between sound apples and bruised apples, the result being a classification rate for sound apples of 84.6% and 77.5% for bruised apples.
• Other works have also demonstrated the value of hyperspectral imaging for the detection of skin defects and damage in other species of fruits like citrus fruits.
• Both simple band ratio algorithms and PCA were tested to discriminate good cucumber skins from those of chilling-injured cucumbers.

5.2. Estimation of internal quality parameters

• Hyperspectral imaging has also been widely used (mostly n apples) to measure internal quality attributes of fruits, such as sugar or SSC, flesh and skin colour, firmness, acidity and starch index, and so forth.
• These results indicated that while LD parameters at single wavelengths were related to fruit firmness, they were insufficient for accurate prediction of fruit firmness.
• The fluorescence and reflectance models both yielded poorer prediction results for TA.
• The starch index was also employed by Nguyen Do Trong et al. (2011) to estimate the optimal cooking time of potatoes.
• Hyperspectral imaging has also been used for determining the internal quality of other fruits and vegetables, apart from apples.

6. Future trends

• The price of the equipment is constantly decreasing, while the technology allows more accurate imaging systems to be developed that are capable of going further into the electromagnetic spectrum.
• This would enable researchers to create new applications oriented towards the non-destructive estimation of internal compounds related with the organoleptic quality or shelf life of the products.
• There are still two challenges to be overcome using this technology.
• In these cases, for example, different studies can obtain different sets of wavelengths for similar applications.

7. Conclusions

• This paper has summarised the current state of the ar on the application of hyperspectral imaging for fruit and vegetable inspection.
• Most of works deal with statistical techniques to reduce the dimensionality of the problem, being the most used based on ANN, PCA, PLS or LDA.
• Many current works try to provide the industry with important practical solutions.
• Very few of them investigate the physical-chemical and biological phenomena that are evidenced in the images.
• In general, this is a technology whose use is beginning to extend to inspect the external and internal quality of many horticultural products, mainly because of the constant price reduction of the components and the increment in computation capacity of modern computers.

Figures

Table 3. Works related with the estimation of internal quality features of fruits

Table 1. Studies related to statistical techniques applied to hyperspectral image processing

Table 2. Works related with the estimation of external quality features of fruits

Full text

Document downloaded from:
This paper must be cited as:
The final publication is available at
Copyright
Additional Information
http://dx.doi.org/10.1007/s11947-011-0725-1
http://hdl.handle.net/10251/68034
Springer Verlag
Lorente, D.; Aleixos Borrás, MN.; Gómez Sanchís, J.; Cubero, S.; García Navarrete, OL.;
Blasco Ivars, J. (2011). Recent advances and applications of hyperspectral imaging for fruit
and vegetable quality assessment. Food and Bioprocess Technology. 5(4):1121-1142.
doi:10.1007/s11947-011-0725-1.

Recent advances in hyperspectral imaging for fruit and vegetable
quality assessment
D. Lorente
1
, N. Aleixos
2
, J. Gómez-Sanchis
3
, S. Cubero
1
, O.L. García-Navarrete
1,4
,
J. Blasco
1
1
Centro de Agroingeniería. Instituto Valenciano de Investigaciones Agrarias. Cra. Moncada-Náquera,
Km. 5, 46113 Moncada, Spain. blasco_josiva@gva.es
2
Instituto Interuniversitario de Investigación en Bioingeniería y Tecnología Orientada al Ser Humano.
Universitat Politècnica de València. Camino de Vera s/n, 46022 Valencia, Spain. naleixos@dig.upv.es
3
Intelligent Data Analysis Laboratory, IDAL. Electronic Engineering Department. Universitat de
València. Dr. Moliner 50, 46100 Burjassot (Valencia), Spain.
4
Departamento de Ingeniería Civil y Agrícola. Universidad Nacional de Colombia - Sede Bogotá. Carrera
30 No. 45-03, Edificio 214, Oficina 206. Bogotá, Colombia
Abstract
Hyperspectral imaging systems are starting to be used as a scientific tool for food quality assessment. A
typical hyperspectral image is composed of a set of a relatively wide range of monochromatic images
corresponding to continuous wavelengths that normally contain redundant information or may exhibit a
high degree of correlation. In addition, computation of the classifiers used to deal with the data obtained
from the images can become excessively complex and time consuming for such high dimensional data
sets and this makes it difficult to incorporate such systems into an industry that demands standard
protocols or high-speed processes. Therefore, recent works have focused on the development of new
systems based on this technology that are capable of analysing quality features that cannot be inspected
using visible imaging. Many of those studies have also centred on finding new statistical techniques to
reduce the hyperspectral images to multispectral ones, which are easier to implement in automatic, non-
destructive systems. This article reviews recent works that use hyperspectral imaging for the inspection of
fruit and vegetables. It explains the different technologies available to acquire the images and their use for
the non-destructive inspection of internal and external features of these products. Particular attention is
paid to the works aimed at reducing the dimensionality of the images, with details of the statistical
techniques most commonly used for this task.
Keywords Computer vision, fruits, vegetables, quality, non-destructive inspection, image analysis,
hyperspectral imaging, multispectral imaging
Nomenclature
2D 2 dimensional
ANN Artificial neural networks
ANOVA Analysis of variance
AOTF Acousto-optic tunable filters
BMP Bitmap image format
CCD Charge-coupled device
FLD Fisher's linear discriminant
FWHM Full width at half maximum
GALDA Genetic algorithm based on LDA
LCTF Liquid crystal tunable filters
LD Lorentzian distribution
LDA Linear discriminant analysis
MC Moisture content
MD Mahalanobis distance
NIR Near-infrared
PCA Principal component analysis
PLS Partial least square
PLSDA PLS discriminant analysis
PLSR PLS regression
RF Radio frequency

RGB Red, Green, Blue colour space
RGBI Red, Green, Blue, Infrared
SAM Spectral angle mapper
SID Spectral information divergence
SSC Soluble solids content
TA Titratable acid
UV Ultraviolet
1. Introduction
The application of machine vision to food analyses has increased considerably in recent years, and has
been used with meat (Du & Sun, 2009), fish (Quevedo et al., 2010; Quevedo & Aguilera, 2010), grains
(Manickavasagan et al., 2010), bread (Farrera-Rebollo et al., 2011), fruits and vegetables (Cubero et al.,
2011), among others. The breadth of applications depends, among many other things, on the fact that
machine vision systems provide substantial information about the nature and attributes of the objects
present in a scene. Another important feature of such systems is that they open up the possibility of
studying these objects in regions of the electromagnetic spectrum where the human eye is unable to
operate, such as in the ultraviolet (UV), near-infrared (NIR) or infrared (IR) regions.
The high risk of human error in classification processes has been underlined and is one of the most
important drawbacks that machine vision can help prevent. In a study carried out with different varieties
of apples, where various shape, size and colour parameters were compared, one of the conclusions
reached was the limited human capacity to reproduce the estimation of quality, which is defined as
inconsistency (Paulus et al., 1997). Moreover, as the number of parameters considered in the decision-
making process increases, so does the rate of error in classification. Furthermore, it should also be
mentioned that automatic inspection allows precise statistics to be generated on aspects related to the
quality of the inspected product, which leads to greater control over it and facilitates its traceability.
In this respect, the quality of a particular fresh or processed fruit or vegetable is defined by a series of
external characteristics that make it more or less attractive to the consumer. Such attributes include its
ripeness, size, weight, shape, colour, the presence of blemishes and disease, the presence or absence of
fruit stems, the presence of seeds, and so on, as well as a series of internal properties like sweetness,
acidity, texture, hardness, etc. that can influence the consumer’s decision as to whether to repeat the
consumption of a particular fruit or not. In sum, they cover all of the factors that exert an influence on the
product’s appearance, on its nutritional and organoleptic qualities or on its suitability for preservation.
Most of these factors have traditionally been assessed by visual inspection or destructive sampling
performed by trained operators, but currently many of them, particularly the external ones, can be
estimated with commercial vision systems (Cubero et al., 2011). These vision systems for fruit sorting are
normally based on colour video cameras that imitate the vision of the human eye by capturing images
using three filters centred on red, green and blue (RGB) wavelengths. Therefore, they are limited to
observing scenes and are usually incapable of obtaining much information about the external or internal
composition of the products.
One way to enhance the capability of traditional colour systems that seek to imitate the human eye is the
use of multispectral systems. A hyperspectral image is composed of a relatively wide range of continuous
wavelengths, whereas a multispectral image consists of a few wavelengths that do not necessarily have to
be continuous. The main advantages of multispectral imaging systems are the relatively low cost of the
system in comparison with hyperspectral systems and the fact that they can be more specific for real
applications. In fact, hyperspectral systems are sometimes used just to select the particular set of
wavelengths that will finally be used. An RGB camera could be considered a particular case of a
multispectral system although it is more common to include wavelengths in frequencies outside the
visible, like NIR. For instance, Aleixos et al. (2002) developed a multispectral camera for the inspection
of citrus fruits which was able to acquire visible and near-infrared images (RGBI) from the same scene.
The same authors also developed specific algorithms for inspecting the size, colour and presence of
defects in citrus at a rate of between 5 and 10 fruits/s. The camera had two CCDs, one of which was a
colour CCD that provided RGB information and the other was monochromatic but coupled to an IR filter,
centred on 750 nm, which provided IR information. For defect detection, a Bayesian discriminant model
was used to segment the images at the pixel level, the independent variables being the grey levels of the
RGBI bands and the classes background, defect and sound skin. The experiments were carried out with
oranges, mandarins and lemons. Comparing results with those obtained using human classification
showed 94% coincidence in the worst case (when the fruit was changing colour from green to orange).

The system was also capable of correctly classifying lemons and mandarins, and detected the external
defects in 93% and 94% of cases, respectively. One of the conclusions was that the B improved the
detection of defects compared to using only RGI, but its contribution was of little importance. Taghizadeh
et al. (2011) compared a conventional RGB imaging system based on a standard still camera with a
hyperspectral imaging system (400-1000 nm) to evaluate the quality of mushrooms by estimating the
hunter L-value, which is the most commonly applied feature for mushroom quality grading. Different
model performance indicators showed the reasonably high potential of hyperspectral imaging models to
predict the L-value for mushroom samples in comparison to RGB-based models.
A multispectral vision system developed by Kleynen et al. (2005) included four wavelength bands in the
visible/NIR range for sorting apples cv. Jonagold based on the presence of defects. They used interference
filters centred at 450, 500, 750 and 800 nm, but since the 500 nm spectral component did not give any
significant information for discriminating between defects and sound tissue, finally this spectral
component was not taken into account in computing the frequency distributions. The 450 nm spectral
band provided significant information with which to identify slight surface defects like russet. The 750
and 800 nm bands, on the other hand, offered good contrast between the defect and the sound tissue and
were well suited to detecting internal tissue damage like hail damage, bruises, and so forth. This system
was recently used by Unay and Gosselin (2006) to study several thresholding and classification-based
techniques for pixel-wise segmentation of multispectral images of cv. Jonagold apples using a multilayer
perceptron-based method in order to detect surface defects. This work was later enhanced by extracting
several features from defective skins with the aim of classifying the fruit in different categories (Unay et
al., 2011).
Bennedsen and Peterson (2005) also developed a multispectral machine vision system with the aim of
detecting surface defects on apples. The system operated on apples that were oriented with the stem/calyx
axis perpendicular to the imaging camera. Images were acquired through two optical filters at 740 and
950 nm. Due to the difference in the detecting ability of the two wavebands used, two training sets were
constructed for each variety: one for 740 nm images and one for 950 nm. In order to evaluate the overall
performance of the system, the binary images resulting from the segmentations of the 740 and 950 nm
images were combined. Apples of eight varieties were used to test the combined performance of the
segmentation routines, with a success rate ranging from 78% to 92%. Ariana et al. (2006) presented an
integrated approach using multispectral imaging in reflectance and fluorescence modes to acquire images
of three varieties of apples (Honeycrisp, Redcort and Red Delicious) to distinguish various defects on
apples, including bitter pit, soft scald, black rot, decay and superficial scald. They acquired eighteen
images from a combination of filters ranging from the visible region through the NIR region and from
three different imaging modes (reflectance -R-, visible light-induced fluorescence and UV-induced
fluorescence) for each apple as a basis for pixel-level classification into normal or damaged tissue. Seven
band pass filters (450, 550, 680, 740, 880, 905, and 940 nm peak transmittance) and a 710 nm long-pass
filter were used in this system.
Blasco et al. (2007) developed a multispectral inspection system to detect and sort citrus fruits according
to 11 different types of external defects by combining the information obtained from four image
acquisition systems that are sensitive to NIR, visible, UV and fluorescence. Compared with the results
obtained using only colour images, the multispectral system showed that the contribution of non-visible
information increased the rate of success in fruit classification by about 78%. This research was later
enhanced in Blasco et al. (2009) to include some morphological features of defects, which raised the
success rate up to 86%, 82% being the success rate when only an RGB camera was used. Single
wavebands can be combined to create spectral indexes. Some of these indexes were studied by Lleó et al.
(2011) to determine the ones that best fit the ripeness prediction of Richlady peaches, with two new
indexes being proposed.
The detection of contaminants can be one of the applications of such multispectral systems. Kalkan et al.
(2011) developed a two-dimensional local discriminant bases algorithm to discriminate between
aflatoxin-contaminated and uncontaminated hazelnuts and red chili peppers flakes. The samples were
screened with 12 different filters, some at 400–510 nm with 10 nm FWHM and others at 550 and 600 nm
with 70 and 40 nm FWHM, respectively. The algorithm classified the flakes into aflatoxin-contaminated
and uncontaminated classes with a 79.2% accuracy rate, so that the level of aflatoxin in the test set was
decreased from 38.26 ppb to 22.85 by removal of the ones that were classified as contaminated. The
hazelnut kernels were independently subjected to two different classifications: first, on the detection of
contamination and, second, on the detection of fungal infestation without considering their aflatoxin

concentrations. A correct classification accuracy of 92.3% was achieved for classifying the hazelnuts as
aflatoxin-contaminated (>4 ppb) or not (<4 ppb).
Internal quality can also be predicted using these systems. For instance, Peng and Lu (2005) presented a
multispectral system with the objective of developing mathematical models to describe the relationship
between fruit firmness and multispectral scattering profiles from apples. Scattering images were acquired
from Red Delicious apples using two different multispectral imaging systems (a rotating filter and a
multispectral imaging spectrograph) at wavelengths of 680 nm, 880 nm, 905 nm and 940 nm with a band-
pass of 10 nm. Each scattering image was reduced to a simple spatial scattering profile through radial
averaging. In a different study aimed at estimating maturity, Lleó et al. (2009) used a multispectral
imaging system to classify peaches into different levels of maturity at harvest and to compare this
classification with reference measurements such as firmness or reflectance at 680 nm achieved with a
visible spectrometer. The proposed system had three band-pass filters centred at 800 nm IR, 675 nm Red
(R) and 450 nm Blue (B), with a bandwidth of 20 nm. Two non-supervised classifications based on the
Ward method were applied on the histograms extracted from the region of interest, i.e. the skin of the
peach. The first classification considered the R channel image of each sample, while the second used the
histograms of the R/IR images, which achieved better results (90% agreement). The use of the R/IR ratio
avoided the effect of fruit shape on light reflectance and thus improved the definition of multispectral
maturity clusters. In contrast, the contribution of the B component in the classification was poor.
Beyond multispectral imaging, the use of hyperspectral sensors makes it possible to conduct a more
sophisticated analysis of the scene by acquiring a set of images corresponding to particular wavelengths,
normally in the visible and NIR part of the electromagnetic spectrum. The reduction in the price of
hyperspectral systems, typically used for remote sensing and meteorology, allows them to be used in
laboratories for food quality and they are an emerging and promising tool for food quality and safety
control, as Gowen et al. (2007) stated in an earlier review of the use of this technology in food inspection.
The acquired multidimensional spectral signature (spectrum) characterising a pixel can be used to analyse
scenes like a standard colour camera but also to obtain information about internal compounds that can be
related with the internal quality of the product.
These systems work with a large number of monochromatic images of the same scene at different
wavelengths, thus enabling simultaneous analysis of the spatial and spectral information. The set of
monochromatic images that are captured constitute a hyperspectral image. As they are made up of a
large collection of images, hyperspectral images constitute a far more extensive source of information
than that provided by a single monochromatic image or a conventional RGB image. The number of
images depends on the spectral resolution of the system used and they are combined by forming a cube
in which two dimensions are spatial (pixels) and the third is the spectrum of each pixel. Without
adequate processing, such a large amount of data, despite being one of the main advantages of
hyperspectral systems, can complicate the extraction of useful information, since much of the
information obtained is redundant or, by nature, cannot be used to distinguish between regions with
similar characteristics (Shaw & Burke, 2003). The demanding industrial restrictions of working in real-
time often make it necessary to reduce the dimensionality of the problem and to select the greatest amount
of non-redundant information from the least number of wavelengths. Unsupervised methods such as
principal component analysis or supervised ones such as linear discriminant analysis are commonly
employed.
Another detail to bear in mind is that when raw hyperspectral images are analysed, it is the radiance of the
scene rather than its reflectance that is being analysed. For these two reasons, when a hyperspectral image
is acquired, first it is necessary to carry out the appropriate compensations in order to separate the
reflectance of the scene from the radiance, and to apply techniques to reduce the amount of information
obtained.
Some techniques for acquiring hyperspectral images e
ven share technology with spectrometry, although
the two techniques should not be confused. Hyperspectral imaging provides spectral and spatial
information (what and where), while spectrometry provides information about spectral information
captured at a particular spot on the sample. To know more about this technology, a good review about
spectrometry was carried out by Nicolaï et al. (2007) or other literature can be consulted (Ozaki et al.,
2006; Sun, 2009). Figure 1 shows a hyperspectral image of an orange with some external defects that are
clear in particular wavelengths and practically invisible in others.

Frequently asked questions

Q1. What are the contributions in "Recent advances in hyperspectral imaging for fruit and vegetable quality assessment" ?

This article reviews recent works that use hyperspectral imaging for the inspection of fruit and vegetables. Particular attention is paid to the works aimed at reducing the dimensionality of the images, with details of the statistical techniques most commonly used for this task. 

Q2. What future works have the authors mentioned in the paper "Recent advances in hyperspectral imaging for fruit and vegetable quality assessment" ?

The future of hyperspectral systems applied to food inspection is promising, since both the industry and consumers are becoming increasing aware of need to ensure the quality and safety of food, and this technology is an important tool for the automatic inspection and monitoring of these parameters. The price of the equipment is constantly decreasing, while the technology allows more accurate imaging systems to be developed that are capable of going further into the electromagnetic spectrum. The partial solution is to search for a small set of important wavelengths that can be used to deal with each problem individually but which sometimes miss important information or limit the potential scope of the final application. 

Q3. What are the essential elements for constructing hyperspectral imaging systems?

The essential elements for constructing hyperspectral imaging systems include light sources, wavelength selection devices, and area detectors (Sun, 2010). 

Q4. What was used to study the changes in reflectance of avocados?

Hyperspectral reflectance imaging was used by Karimi et al. (2009) to study the changes in reflectance (350-2500 nm) of avocados coated with different formulations. 

Q5. What is the common use of spectral imaging in the inspection of fruits and vegetables?

Detection of skin defects is one of the most widespread uses of hyperspectral imaging in the inspection of fruits and vegetables, since the perceived quality is highly associated with a good appearance of the product. 

Q6. How many varieties of apples were used to test the combined performance of the segmentation routines?

Apples of eight varieties were used to test the combined performance of the segmentation routines, with a success rate ranging from 78% to 92%. 

Q7. How many wavelengths were found that could be used to detect bruises in apples?

Three wavelengths in the NIR region (750, 820, 960 nm) were found that could potentially be implemented in multispectral imaging systems for the detection of bruises in this cultivar of apples. 

Q8. How many wavelengths were required to predict the maturity of banana fruits?

Eight wavelengths were required to predict the maturity stages of banana fruits representing the quality attribute in terms of the features that were studied. 

Q9. Why did the algorithm adapt to the large variability of intensities and shapes of the image regions?

Due to the unsupervised nature of the procedure, it could adapt itself to the large variability of intensities and shapes of the image regions. 

Q10. What is the method for detecting surface defects in apple inspection systems?

Results showed that among many classification and thresholding-based methods, MLP was the most promising for segmenting surface defects in high-speed machine vision-based apple inspection systems. 

Q11. What is the way to assess the internal quality of apples?

Other applications of hyperspectral imaging systems to the assessment of the internal quality of apples have been cited in recent literature, such as chilling injury detection. 

Q12. How did Polder et al. (2004) study ripening in tomatoes?

In order to study ripening in tomatoes, Polder et al. (2004) analysed concentrations of different compounds using HPLC and by analysing spectral images using PLS regression at the pixel level and at the tomato level. 

Q13. How many spectral ranges were used to determine the sugar content of apples?

On applying the PLS method to the spectral profiles of the fruits, it was found that the optimal spectral range for sugar content was 704-805 nm. 

Q14. What are the attributes of a fruit that can influence the decision to eat it?

Such attributes include its ripeness, size, weight, shape, colour, the presence of blemishes and disease, the presence or absence of fruit stems, the presence of seeds, and so on, as well as a series of internal properties like sweetness, acidity, texture, hardness, etc. that can influence the consumer’s decision as to whether to repeat the consumption of a particular fruit or not. 

Q15. What are the common techniques used to reduce the dimensionality of the problem?

Most of works deal with statistical techniques to reduce the dimensionality of the problem, being the most used based on ANN, PCA, PLS or LDA. 

Q16. How was the accuracy of the classification of aflatoxin-contaminated hazelnuts?

The algorithm classified the flakes into aflatoxin-contaminated and uncontaminated classes with a 79.2% accuracy rate, so that the level of aflatoxin in the test set was decreased from 38.26 ppb to 22.85 by removal of the ones that were classified as contaminated. 

Q17. What was the method for comparing LD parameters to fruit firmness?

In addition, for each wavelength, a multi-linear regression analysis was attempted between firmness and LD parameters for both cultivars.