hyperspectral imaging for food quality analysis and control || visualization of sugar distribution...

CHAPTER 11

Hyperspectral Imaging for Food Quality Analysis an

Copyright � 2010 Elsevier Inc. All rights of reproducti

Visualization of SugarDistribution of Melons byHyperspectral Technique

Junichi Sugiyama, Mizuki TsutaNational Food Research Institute, Tsukuba, Ibaraki Japan

CONTENTS

Introduction

Visualization by VisibleWavelength Region

Visualization by SugarAbsorption Wavelength

Conclusions

Nomenclature

References

11.1. INTRODUCTION

In Japan, automated sweetness sorting machines for peaches, apples, and

melons based on near-infrared (NIR) spectroscopy techniques have been

developed and are now in use in more than 172 packing houses (Hasegawa,

2000). However, parts of a fruit sorted by the machine as sweet may

sometimes taste insipid because of an uneven distribution of the sugar

content. Visualization of the sugar content of a melon is expected to be

useful not only for evaluation of its quality but also for physiological

analysis of the ripeness of a melon. There have been several attempts to

obtain a distribution map of the constituents of agricultural produce

(Bertrand et al., 1996; Ishigami & Matsuura, 1993; Robert et al., 1991,

1992; Taylor & McClure, 1989). However, a quantitatively labeled distri-

bution map has not yet been obtained.

On the other hand, recently, device and personal computer (PC)

technology have advanced greatly. Cooled charge-coupled device (CCD)

imaging cameras with a wide dynamic range have been introduced, which

makes quantitative measurements possible. Modern PCs can easily accept

and/or process a large volume of data. Taking advantage of these, the

conventional NIR spectroscopy technique, which is a technique for point

measurement, could be extended to two-dimensional measurements. This

chapter thus discusses the development of a technique for visualization of

the sugar content of a melon by applying NIR spectroscopy to each pixel

in an image.

d Control

on in any form reserved. 349

CHAPTER 11 : Visualization of Sugar Distribution of Melons by Hyperspectral Technique350

11.2. VISUALIZATION BY VISIBLE WAVELENGTH

REGION

11.2.1. Melons

Maturity levels of melons at which they are harvested significantly affect

their sugar content distribution. Sugiyama (1999) used a hyperspectral

technique to compare three ripeness stages (unripe, mature, and fully

mature) of Andes melons harvested in Tsuruoka, Yamagata Prefecture, Japan,

in 1998. Unripe melons were harvested 6 days earlier, and fully mature ones

5 days later than the mature melons. The mature melons were harvested 55

days after pollination. Two melons at each stage, that is, a total of six melons,

were investigated. There were cracks on the bottom of the fully mature

melons because of overripening. Each sample was sent to the laboratory the

day after the harvest using a special delivery service. Experiments were

carried out in a dark room at 25 �C.

11.2.2. NIR Spectroscopy

11.2.2.1. Measurement of spectra and sugar content

In order to determine the wavelength that has a high correlation with sugar

content, spectra between 400 and 1100 nm in the flesh of a melon were

analyzed using a NIR spectrometer (NIRS 6500, FOSS NIR Systems, Silver

Spring, MD, USA). This wavelength range covers the spectrum of the CCD

camera used for the imaging application. A cylindrical sample with a diameter

of 20 mm was extracted from the equator of the melon by a stainless steel

cylinder with a knife edge at one end (Figure 11.1). A spectrum of the sample’s

inner surface was obtained using a fiber-type detector with an interactance

mode (Kawano et al., 1992). The wavelength interval was 2 nm and the

number of scans was 50. Then, the measured portion was cut into a 2 mm-

thick slice with a kitchen knife and squeezed with a portable garlic crusher

to measure the Brix sugar content using a digital refractometer (PR-100,

ATAGO, Yorii, Saitama, Japan). The measurements of the spectrum and the

sugar content were repeated similarly at various depths within the melon.

11.2.2.2. Wavelength selection by NIR spectroscopy

Figure 11.2 shows the simple correlation coefficient calculated by a standard

regression model with no data pretreatment between the absorbance at each

wavelength and the sugar content. Each line was calculated from 22 slices

made from two cylindrical samples which were extracted from a melon. It is

clear that the wavelength of 676 nm exhibits the maximum absolute

Repeat

2 mm sliced

Squeezed

Refractometer(Brix)

NIR spectrometer

FIGURE 11.1 Sample preparation for measurements of NIR spectra and the sugar

content

1.0

0.8 Earl’s melonAndes melon

676nm

Wavelength [nm]

0.6

0.40.2

0

−0.2

Co

rrelatio

n co

efficien

t

−0.4

−0.6−0.8−1.0

400 500 600 700 800 900 1000 1100

FIGURE 11.2 Correlation coefficients at each wavelength between the absorbance

and the Brix sugar content

Visualization by Visible Wavelength Region 351

correlation, although it is inversely correlated with the sugar content for both

the Andes and Earl’s varieties of melon. Because of the inverse correlation, it

seems that 676 nm is not a direct absorption band of sugar but a wavelength of

secondary correlation (Osborne & Feam, 1986) with a component inversely

proportional to the sugar content. It is actually close to the absorption band of

chlorophyll (Nussberger et al., 1994; Watada et al., 1976) and there are some

implications for other produce (Izumi et al., 1990; Morita et al., 1992). With

the fact that absorbance at 676 nm is inversely correlated with the sugar

content for its visualization, the interpretation of 676 nm is also important in

terms of the physiological aspects and must be further studied.

Adapter

Camera lens

Interference filter

Quartz glass

Moving wall

Antireflection velvet sheet

Sample

Iron bench

Optical fiber illuminator

To CCD controllerand computer

CC

D c

amer

a

FIGURE 11.3 Configuration of an apparatus for spectroscopic image acquisition


11.2.3. Imaging Spectroscopy

11.2.3.1. Instrumentation

Figure 11.3 shows the configuration of the imaging apparatus used to obtain

the spectroscopic images. Although a monochrome CCD camera normally

has an 8-bit (256 steps) analog-to-digital (A/D) resolution, a cooled CCD

camera (CV-04II, Mutoh Industries Ltd., Tokyo, Japan) with a 16-bit (65 536

steps) A/D resolution was adopted. The advantage of the high A/D resolution

is that each pixel can function as a detector of the NIR spectrometer for

quantitative analysis. The CCD camera has a linear intensity characteristic

(g ¼ 1) and no antiblooming gate for quantitative analysis. To decrease the

electrical dark current noise of the CCD camera, both double-stage ther-

moelectric cooling and water cooling were utilized. A camera lens (FD28 mm

F3.5 S.C., Canon, Tokyo, Japan) with an interference filter (J43192, Edmond

Scientific, Tokyo, Japan) was installed through the camera adapter (Koueisya,

Kawagoe, Saitama, Japan). The interference filter had band-pass character-

istics of 676 nm at the central wavelength, which was determined in the NIR

spectroscopic experiment (see 11.2.2.2), and 10 nm at half-bandwidth. The

illuminator (LA-150S, Hayashi Watch-Works, Tokyo, Japan) had a tungsten–

halogen bulb driven by direct current to reduce optical noise. The source light

was introduced into two fiber-optic probes, illuminating a sample from two

different positions so as not to create any shadows or direct reflection.

A sample was placed perpendicularly on the quartz glass maintaining

ncbi-n:J43192


a constant focal distance between the CCD camera and the sample

(Figure 11.3). The sample was supported by the moving wall covered with

a black antireflection velvet sheet.

11.2.3.2. Image of half-cut melon for sugar distribution map

Each melon (six in total) was cut vertically in half with a kitchen knife.

Spectroscopic images of the surface of a half-cut melon at 676 nm were taken

with an aperture of 16 (F16) and an exposure period of 0.5 seconds. The

cooling temperature of the CCD camera was �15 �C. The size of the image

was 768� 512 pixels. After obtaining a vertical image of the half-cut melon,

the melon was cut in a horizontal plane, and a horizontal image of the

quarter-cut melon was captured under the same conditions as described

above.

11.2.3.3. Partial image for sugar content calibration

After obtaining the aforementioned images, two cylindrical samples with

a diameter of 20 mm were extracted from the equator of the same melon. In

the same manner as in the NIR spectroscopic experiment (Figure 11.1), an

image of the surface was taken at 676 nm using the CCD camera under the

same conditions as for the half-cut melon described previously. Then

a 2 mm-thick slice was cut off and squeezed for the measurement of sugar

content. These procedures were repeated until the rind appeared.

11.2.3.4. Noise corrections

Images acquired using a CCD camera include (i) thermal noise due to dark

current thermal electrons, (ii) bias signals to offset the CCD slightly above

zero A/D counts, (iii) sensitivity variations from pixel to pixel on the CCD,

and (iv) lighting variations on the sample’s surface. In order to compensate

for the above, the following noise corrections (Fukushima, 1996; Morita

et al., 1992; SBIG, 1998) were carried out:

Processed image ¼ raw image� dark frame

flat field� dark frame of flat field�M (11.1)

In Equation (11.1), the dark frame is the image acquired under the same

conditions as the raw image except for the absence of lighting. Subtracting it

from the raw image allows corrections for (i) thermal noise and (ii) bias

signals. On the other hand, the flat field is obtained by taking an exposure of

a uniformly lit ‘‘flat field’’ such as a Teflon board. After subtracting the dark

frame of the flat field, in the same way as the numerator, the ratio between

the two images is compensated for the effect of (iii) sensitivity variations and


(iv) lighting variations. M is the intensity value averaged over all pixels of the

flat field after dark frame subtraction (¼ denominator in Equation 11.1). The

multiplier M restores the ratio of the images to the image intensity level.

All of these image processes were carried out using software (CCD Master,

Mutoh Industries Ltd., Tokyo, Japan) compatible with the CCD camera.

11.2.3.5. Conversion from intensity into sugar content

Each pixel of the image processed using Equation (11.1) has 16 bits, that is,

65 536 levels of intensity. The method for converting the intensity into sugar

content on an image data was developed in accordance with NIR spectros-

copy. Based on the fact that the functional group of chemical compounds

responds to near-infrared radiation, NIR spectroscopy can measure the

amount of a specific constituent from its absorbance at several wavelengths

(Osborne & Feam, 1986). Absorbance A can be defined as follows:

A ¼ logðIs=IÞ (11.2)

where Is is the intensity of reflection of a white standard board; I is the

intensity of reflection of the sample.

Because Is and I correspond to the denominator and the numerator in the

first term of Equation (11.1), respectively, Equations (11.3) and (11.4) are

introduced.

Is ¼ flat field� dark frame of flat field (11.3)

I ¼ raw image� dark frame (11.4)

Considering Equations (11.1 to 11.4), absorbance A in the spectroscopic

image can be expressed as follows:

A ¼ logðIs=IÞ¼ logðM=processed imageÞ¼ logðM=RÞ

(11.5)

where R is the intensity of the reflection of each pixel in a processed image,

and M is the average intensity of reflection of the flat field.

On the other hand, the NIR spectroscopic experiment indicated that the

absorbance at 676 nm was correlated with the sugar content. The same

relationship in the image system of this experiment could be confirmed by

using the following procedure: (i) calculation of the average intensity of

a partial image of 20 mm diameter, (ii) conversion of the average intensity

into absorbance using Equation (5), and (iii) plotting the relationship between

18

16

14

12

10

8r=0.995y=−15.799x+18.029

r=0.976y=−27.08x+28.556

y=−23.723x+22.473

UnripeMatureFully mature

r=0.982

6

40.3 0.4 0.5 0.6

Absorbance log (M/R)

Su

gar co

nten

t [B

rix]

0.7 0.8 0.9

FIGURE 11.4 Calibration curves between the sugar content and the absorbance at

676 nm by the imaging system


the absorbance and sugar content for each partial image. A total of six

melons, two for each stage of ripeness, were analyzed, and the representative

results for unripe, mature, and fully mature melons are shown in Figure 11.4.

The number of symbols in Figure 11.4 indicates the number of the sliced

samples subjected to measurements of the absorbance and sugar content.

Each calibration curve is slightly different from the others because the light

condition had been adjusted for each sample to avoid direct reflection (glit-

tering) on rugged portions. This adjustment changed the lighting intensity

level, which is not corrected by Equation (11.1), and subsequently affected

the calibration curves. However, it was confirmed that the image system can

reveal the sugar content using the calibration curves for each sample.

11.2.4. Visualization of Sugar Distribution

The image of a half-cut melon for drawing a sugar distribution map was

corrected for noise using Equation (11.1). The processed image was con-

verted into an absorbance image using Equation (11.5). Then, an image of

sugar content was calculated by applying the calibration curve to each pixel of

the absorbance image. These actual procedures, from retrieval of the pro-

cessed image to saving the sugar content image, were carried out using an

original program written in Visual Basic (Microsoft, Redmond, WA, USA).

Finally, the sugar content image was visualized with a linear color scale by the

visualization software (AVS/Express Viz, Advanced Visual Systems,

Waltham, MA, USA). Figure 11.5 shows the results of visualization of the

sugar content corresponding to unripe, mature, and fully mature melons,

FIGURE 11.5 Sugar distribution map for unripe, mature, and fully mature melons.

(Full color version available on http://www.elsevierdirect.com/companions/

9780123747532/)


respectively. Since the measurements were carried out just after the harvest,

the flesh of each melon was sufficiently hard that it was actually difficult to

tell the differences in sugar content by the naked eye. However, as a result of

the visualization, the sugar content at each stage of ripeness was clarified. In

particular, in the mature and fully mature melons the distribution of the

sugar content varies among different parts of the fruit, indicating the

importance of the part of the fruit sampled in the conventional measurement

of the sugar content with a refractometer. In addition, as shown in the mature

melon, the upper part had higher sugar content than the bottom part. These

results suggest that the visualization technique by NIR imaging could

become a useful new method for quality evaluation of melons. Moreover,

there is a good possibility that applying several wavelengths for the calibra-

tion curve could allow visualization of many more constituents of other

agricultural products.

11.3. VISUALIZATION BY SUGAR ABSORPTION

WAVELENGTH

The former method cannot be applied to a red-flesh melon because it

depends not on the absorption band of the sugar, but on the color information

http://www.elsevierdirect.com/companions/9780123747532/


Visualization by Sugar Absorption Wavelength 357

at 676 nm. Therefore, Tsuta et al. (2002) developed a universal method for

visualization of sugar content based on the absorption band of the sugar in

the NIR wavelength region.

11.3.1. Melons

Two green-flesh melons (Delicy) and three red-flesh melons (Quincy) were

prepared for NIR spectroscopy and another red-flesh melon (Quincy) for

imaging. They were obtained from a store and left overnight in a dark room at

25 �C before the experiment. The experiments were carried out in the same

room.

11.3.2. NIR Spectroscopy for Sugar Absorption Band

11.3.2.1. Measurement of spectra and sugar content

To specify the absorption band of sugar, a NIR spectrometer (NIRS 6500,

FOSS NIRSystems, Silver Spring, MD, USA) and a digital refractometer (PR-

100, ATAGO, Yorii, Saitama, Japan) were utilized (Figure 11.6). Pretreatment

of the acquired spectra and a multi-linear regression (MLR) analysis were

carried out using spectral analysis software (VISION, FOSS NIRSystems,

Silver Spring, MD, USA).

A 25 mm-diameter cylindrical sample (Figure 11.6a) was extracted from

the equator of a melon using a stainless steel cylinder with a knife edge at one

end. A spectrum of the sample’s inner surface was obtained using a fiber-

optic probe (Figure 11.6b) of the NIR spectrometer in the interactance mode

25 mm

ExtractedSliced

Repeated

Microtube (d)

(c)(a)

(b)

(e)

Fiber-optic Probe

Supernatantjuice

Refractometer(Brix)

Centrifuged

NIR spectrometer(NIRS 6500)

Frozen and defrosted

FIGURE 11.6 NIR spectroscopy for evaluation of sugar content of melons


(Kawano et al., 1992). The wavelength interval was 2 nm, and the number of

scans was 50. The measured portion was then cut into a 1 mm-thick slice

(Figure 11.6c) using a handicraft cutter and put into a 1.5 ml microtube

(Figure 11.6d) to be frozen and defrosted. This process was intended to break

the cell walls of the portion in order to extract a sufficient amount of juice for

measuring the sugar content (Martinsen & Schaare, 1998). The portion was

then centrifuged for 10 min at 10 000 rpm to extract juice. The �Brix sugar

content of the juicewas measured using the digital refractometer (Figure 11. 6e).

A set of the spectrum and the sugar content measurements for every

1 mm-thick slice was repeated from the inner surface toward the rind. Each

raw spectrum was converted into a second-derivative spectrum to decrease the

effect of spectral baseline shifts (Iwamoto et al., 1994; Katsumoto et al.,

2001). An MLR analysis was carried out for all of the data sets to acquire the

calibration curve for the sugar content and the second-derivative spectra.

11.3.2.2. Calculation of second-derivative spectrum

Derivative methods are important pretreatment methods in NIR spectros-

copy. The second-derivative method is most often used because of its

following merits (Iwamoto et al., 1994; Katsumoto et al., 2001):

1. Positive peaks in a raw spectrum are converted into negative peaks in

a second-derivative spectrum.

2. The resolution is enhanced for the separation of overlapping peaks

and the emphasis of small peaks.

3. The additive and multiplicative baseline shifts in a raw spectrum are

removed.

By applying the truncated Taylor series expansion, a second-derivative

spectrum can be calculated as follows (Morimoto et al., 2001):

f2ðxÞ ¼ fðxþ DxÞ � 2� fðxÞ þ fðx� DxÞDx2

(11.6)

where f(x) is the spectral function at x and f 2(x) is the second-derivative

function at x. Actual spectral data, however, take discrete values because of

the limited wavelength resolution of NIR spectrometers. Therefore,

a second-derivative spectrum is calculated as follows in NIR spectroscopy

(Katsumoto et al., 2001):

d2Ai ¼ Aiþk � 2� Ai þ Ai�k (11.7)


where Ai is an absorbance at i nm, d2Ai is a second-derivative absorbance at

i nm, and k is a distance between the neighboring wavelengths, which is

called a derivative gap. Equation (11.7) shows that absorbances at three

wavelengths of i, iþ k and i� k are enough for calculating the second-

derivative absorbance at i nm. It also indicates that the imaging system can

acquire a second-derivative spectroscopic image using three band-pass filters.

11.3.2.3. Absorption band of sugar by NIR spectroscopy

One hundred and fifty-seven spectra were obtained as a result of NIR spec-

troscopy, and the MLR analysis of the spectra revealed that the second-

derivative absorbances at 874 and 902 nm were highly correlated with the

sugar content as shown in Table 11.1. The correlation was maintained at

a high level of more than 0.99, whereas the derivative gap changed from 20 to

36 nm. The derivative gap was selected to decrease the number of band-pass

filters for the imaging system. Conventionally, six band-pass filters are

necessary to acquire two second-derivative spectroscopic images. However,

when the derivative gap of 28 nm was adopted, only four band-pass filters,

that is, 846, 874, 902, and 930 nm, were sufficient for the analysis. This is

because 874 and 902 nm overlapped between two second derivatives (indi-

cated by bold italic digits in Table 11.1). When 28 nm was selected as the

derivative gap, the calibration curve was as follows:

�Brix ¼ 21:93� 410:76 d2A902 þ 1534:76 d2A874 (11.8)

Table 11.1 Relationship among the derivative gap, correlation, and necessaryband-pass filters

Necessary band-pass filters (nm)

Gap (nm) R For d 2A874 For d 2A902

4 0.976 870 874 878 898 902 906

8 0.975 866 874 882 894 902 910

12 0.983 862 874 886 890 902 914

16 0.988 858 874 890 886 902 918

20 0.990 854 874 894 882 902 922

24 0.991 850 874 898 878 902 926

28 0.991 846 874 902 874 902 930

32 0.991 842 874 906 870 902 934

36 0.990 838 874 910 866 902 938

40 0.988 834 874 914 862 902 942

Note: Bold type denotes overlapping wavelengths.

17.0

14.8

12.6

10.4

8.2

6.0 6.0 8.2 10.4 12.6 14.8

R = 0.991SEC = 0.333

17.0Actual °Brix value

Calcu

lated

°B

rix valu

e

FIGURE 11.7 Calibration by NIR spectroscopy


The curve had a high correlation with the sugar content (R ¼ 0.991), and

the standard error of calibration was 0.333 (Figure 11.7). The second-

derivative absorbance at 902 nm had an inverse correlation with the sugar

content, which indicated that the raw absorbance had a positive correlation

with it. In addition, several publications (Ito et al., 1996; Kawano & Abe,

1995; Kawano et al., 1992, 1993; Temma et al., 1999) indicated that 902 nm

is one of the typical absorption bands of sugar components. On the other

hand, 874 nm can be considered as the reference wavelength to compensate

for different surface conditions or some other influences. As a result, four

band-pass filters of 846, 874, 902, and 930 nm were adopted for the imaging

system.

11.3.3. Imaging Spectroscopy for Sugar Absorption Band

11.3.3.1. Instrumentation

Figure 11.8 shows the configuration of the apparatus for obtaining spec-

troscopic images. The cooled CCD camera (CV-04 II, Mutoh Industries

Ltd., Tokyo, Japan) had a 16-bit (65 536 steps) A/D resolution, a linear

intensity characteristic (g ¼ 1), and no antiblooming gate, so that each

pixel could function as a detector of an NIR spectrometer for quantitative

analysis. To decrease the electrical dark current noise of the CCD camera,

both double-stage thermoelectric cooling and water cooling were utilized.

A filter adapter with a filter holder (Koueisha, Kawagoe, Saitama, Japan)

and a camera lens (FD28 mm F3.5 S. C., Canon, Tokyo, Japan) were

installed in the CCD camera. The filter holder had four holes to which four

filters could be fitted. The four filters in this experiment (BWEx; x ¼ 846,

Near-infrared illuminator

CCD camera

Adapter Sample

Calibration sample

Iron bench

Line-shaped light guides

Near-infrared illuminator

Camera lens

Band-pass filters

To CCD controller and computer

FIGURE 11.8 Imaging system for spectroscopic image acquisition at different

wavelengths


874, 902, 930, Koshin Kogaku Filters, Hatano, Kanagawa, Japan) were

designed to have band-pass characteristics of x nm at the central wave-

length, and their details are shown in Table 11.2. The wavelengths of 902

and 874 nm were determined in the NIR spectroscopic experiment (see

11.3.2.3), and the others were their neighboring wavelengths selected to

calculate the second-derivative absorbances. The near-infrared illuminator

(LA-100IR, Hayashi Watch-Works, Tokyo, Japan) irradiated only NIR light

because a NIR reflecting mirror was installed around a tungsten–halogen

bulb, and a high-pass filter, which transmits only light above 800 nm, was

attached to the irradiation hole. The source light was introduced into line-

shaped light guides through a fiber-optic probe, illuminating a sample from

two different positions in order not to create any shadows or direct

reflection. Previously, a quartz glass had been placed on the surface of the

sample to maintain a constant focal distance between the CCD camera and

the sample (Sugiyama, 1999). In this experiment, however, a direct

reflection image of the CCD camera on the glass was observed because the

intensities of the sample and unnecessary images were both low; these

were enhanced by a long exposure period. Therefore, the quartz glass was

not adopted in this experiment; instead, the sample was placed on an iron

bench facing the camera.

Table 11.2 Characteristics of the band-pass filters

Central wavelength (nm)

Model Specified value Measured value Bandwidth (nm)

BWE846430 846.0 � 2.0 847.5 13.3

BWE874430 874.0 � 2.0 875.8 13.6

BWE902430 902.0 � 2.0 900.0 16.0

BWE930430 930.0 � 2.0 928.5 16.0


11.3.3.2. Acquisition of the spectroscopic images

Using the imaging system, whole images of the surface of a half-cut sample

were taken at 846, 874, 902, and 930 nm at an exposure period of 3.7 s.

The binning mode of 2� 2, that is, four pixels of the CCD camera were

combined to function as one pixel, was applied to acquire a higher sensi-

tivity (Fukushima, 1996). The temperature of the CCD camera was

maintained at �20 �C. The size of the image was 384� 256 pixels after

binning. After a half-cut image had been captured, two 25 mm-diameter

cylindrical samples were extracted from the equator of the same melon.

These cylindrical samples were used to acquire a sugar content calibration

curve based on the imaging system. In the same manner as in the case of

a half-cut sample, images of the surface of the cylindrical samples were

taken, after which a 1 mm-thick slice was obtained and the �Brix sugar

content was measured as described for the NIR spectroscopic experiment

(see Figure 11.6). Image capture and measurement of the sugar content

were repeated until the rind appeared.

11.3.3.3. Image processing for calibration

The obtained raw images of the cylindrical samples include (1) thermal

noise, (2) bias signals to offset the CCD slightly above zero A/D counts,

(3) sensitivity variations from pixel to pixel on the CCD, and (4) lighting

variations on the sample’s surface (Fukushima, 1996). To compensate for the

above effects, noise and shading corrections were carried out for all images.

The average intensity of the images of the cylindrical sample was converted

into the average absorbance based on the spectroscopy theory (Figure 11.9).

These processes were described in Section 11.2.3.5. Once the average

absorbance at each wavelength was obtained, the second-derivative absor-

bances at 902 and 874 nm were calculated as follows (Katsumoto et al., 2001;

Morimoto et al., 2001) (see 11.3.2.2):

846 nm 874 nm

I 846

A 846

I 874

A 874

A 874 A 902

Brix = a.d2A

874+b.d

2A

902+c

d2

d2

I 902

A 902

I 930

A 930

902 nm 930 nm

Average intensity

Noise/shading correction

Calculation of the average

intensity within the circle

Conversion:

Differential Calculus

(Eqs 11.9 and 11.10)

MLR Analysis:

Intensity

Two d2

A vs. Actual Brix

Absorbance

Spectroscopic images of

cylindrical samples

Corrected images

Average absorbance

2 derivative absorbancend

Calibration curve

— ———

— —

— —

— —

Images and data Procedure

FIGURE 11.9 Image processing procedure for calibration


d2A902 ¼ A930 � 2� A902 � A874 (11.9)

d2A874 ¼ A902 � 2� A874 � A846 (11.10)

where Ai is the absorbance at i nm and d2Ai is the second-derivative absor-

bance at i nm. Then, MLR analysis using these second-derivative absor-

bances was carried out to acquire the calibration curve for the sugar content

of the imaging system.

11.3.3.4. Calibration by the imaging system

A calibration curve for the second-derivative absorbance and the sugar

content of the imaging system was obtained by MLR analysis of 33 slices

from the cylindrical samples (Figure 11.10), which is given below:

�Brix ¼ 19:01� 438:84 d2A902 þ 70:32 d2A874 (11.11)

The second-derivative absorbance at 902 nm had an inverse correlation

with the sugar content in Equation (11.11), which is the same as in Equation

(11.8). Equation (11.11) had a high correlation of R ¼ 0.891, and the stan-

dard error of calibration was 1.090. Therefore it can be considered that the

imaging system adopted has sufficient capability to visualize the sugar

content.

14.0

12.0

10.0

8.0

6.06.0 8.0

R = 0.891SEC = 1.090

12.010.0 14.0

Calcu

lated

°B

rix v

alu

e

Actual °Brix value

FIGURE 11.10 Calibration for absorbance and the sugar content by imaging


11.3.4. Visualization of Sugar Distribution by Its

Absorption Band

The intensity of each pixel on the half-cut sample image was converted into

the second-derivative absorbances at 902 and 874 nm in the same manner

as in the case of cylindrical samples (Figure 11.11). The acquired calibration

curve was applied to these two second-derivative absorbances at each pixel

in order to calculate the sugar content. The sugar content was then visu-

alized by mapping the value with a linear color scale. Original image

Images and data Procedure

Apply a calibraton

curve (Eq 11.11)

Color mapping

d2A

874

Apply to

each pixel

d2A

902

Brix = a.d2A

874+b.d

2A

902+c

846 nm 874 nm 902 nm 930 nmSpectroscopic images of the

half-cut sample

2nd

derivative absorbance

Sugar content

Sugar distribution map

Image processing procedure (Figure 11.9)

FIGURE 11.11 Visualization procedure. (Full color version available on http://www.

elsevierdirect.com/companions/9780123747532/)



FIGURE 11.12 Sugar distribution map of a half-cut red-flesh melon. (Full color version

available on http://www.elsevierdirect.com/companions/9780123747532/)

Conclusions 365

processing software was utilized to process images and to construct a sugar

distribution map.

A sugar distribution map of the half-cut red-flesh melon was constructed

by applying Equation (11.11) to each pixel of the processed images

(Figure 11.11). In Figure 11.12, sugar contents ranging from 2 to 18 �Brix

were assigned a linear color scale. The color changes gradually from blue to

red as the sugar content increases. Although it was difficult to differentiate

the sugar distribution by the naked eye, Figure 11.12 shows that the sugar

content increases from the rind to near the seeds. It also indicates that the

central upper part of the sample was sweeter than the bottom part, which is

the reverse of the general notion in Japan. These results suggest that NIR

imaging could become a useful method for evaluating the distribution of

sugar in melons. In addition, further studies may lead to the application of

this method not only to various varieties of melons but also to other

constituents of other agricultural products because it does not depend on

color information.

11.4. CONCLUSIONS

The relationship between the sugar content and absorption spectra can be

investigated by using a near infrared (NIR) spectrometer to visualize the

sugar content of a melon. The absorbance at 676 nm, which is close to the

absorption band of chlorophyll, exhibited a strong inverse correlation with



the sugar content. A high-resolution cooled charge-coupled device (CCD)

imaging camera fitted with a band-pass filter of 676 nm was used to capture

the spectral absorption image. The calibration method was used for con-

verting the absorbance values on the image into the �Brix sugar content in

accordance with NIR spectroscopy techniques. When this method was

applied to each pixel of the absorption image, a color distribution map of the

sugar content could be constructed.

In addition, a method for visualizing the sugar content based on the

sugar absorption band was also developed. This method can avoid bias

caused by the color information of a sample. NIR spectroscopic analysis

revealed that each of the two second-derivative absorbances at 874 and

902 nm had a high correlation with the sugar content of melons. A high-

resolution cooled CCD camera with band-pass filters, which included the

above two wavelengths, was used to capture the spectral absorption image

of a half-cut melon. A color distribution map of the sugar content on the

surface of the melon was constructed by applying the NIR spectroscopy

theory to each pixel of the acquired images. As a result, NIR spectroscopy

theory can be extended to imaging applications with a high resolution CCD

camera. Constructing the calibration method by the imaging system is the

key point of this method because it is impossible to measure the actual

sugar content of each pixel. Because an indium gallium arsenide (InGaAs)

camera that can detect longer wavelengths (900–1600 nm) is available

nowadays, wider applications can be expected using hyperspectral imaging

techniques.

NOMENCLATURE

Symbols

A absorbance

Ai absorbance at i nm

d2Ai second-derivative absorbance at i nm

f(x) spectral function at wavelength of x

f2(x) second-derivative spectral function at wavelength of x

I intensity of reflection of a sample

Is intensity of reflection of a white standard board

k derivative gap

M average intensity value of pixels of flat field after dark frame

subtraction

R intensity of the reflection of each pixel in a processed image

References 367

Abbreviation

A/D analog-to-digital

CCD charge-coupled device

InGaAs indium gallium arsenide

NIR near-infrared

PC personal computer

REFERENCES

Bertrand, D., Robert, P., Novales, B., & Devaux, M. (1996). Chemometrics ofmultichannel imaging. In A. M. C. Davies, & P. Williams (Eds.), Near infraredspectroscopy: the future waves (pp. 174–178). Chichester, UK: NIRPublications.

Fukushima, H. (1996). How to use the CCD camera. In Reikyaku CCD nyuumon(pp. 73–132). Tokyo: Seibundou Sinkousha.

Hasegawa, Y. (2000). Merits and demerits of the automated sweetness sortingtechniques. Fresh Food System, 30, 74–77.

Ishigami, K., & Matsuura, H. (1993). Studies on the sugar distribution andcomposition of muskmelon fruit. Bulletin of Shizuoka Agricultural Experi-ment Station, 37, 33–40.

Ito., M., Iida, J., Terashima, A., & Kishimoto, T. (1996). Non-destructive sugarcontent measuring method (Japanese Patent). JP 08–327536 A.

Iwamoto, M., Kawano, S., & Uozumi, J. (1994). Data processing method. In Kin-Sekigai Bunkouhou Nyuumon (pp. 62–95). Tokyo, Japan: Saiwai Shobou.

Izumi, H., Ito, T., & Yoshida, Y. (1990). Seasonal changes in ascorbic acid, sugarand chlorophyll contents in sun and shade leaves of satsuma mandarin andtheir interrelationships. Journal of the Japanese Society for HorticulturalScience, 59, 389–397.

Katsumoto, Y., Jiang, J., Berry, R. J., & Ozaki, Y. (2001). Modern pretreatmentmethods in NIR spectroscopy. Near Infrared Analysis, 2, 29–36.

Kawano, S., & Abe, H. (1995). Development of a calibration equation withtemperature compensation for determining the Brix value in intact peaches.Journal of Near Infrared Spectroscopy, 3, 211–218.

Kawano, S., Fujiwara, T., & Iwamoto, M. (1993). Nondestructive determin-ation of sugar content in satsuma mandarin using near infrared (NIR)transmittance. Journal of the Japanese Society for Horticultural Science, 62,465–470.

Kawano, S., Watanabe, H., & Iwamoto, M. (1992). Determination of sugarcontent in intact peaches by near infrared spectroscopy with fiber optics ininteractance mode. Journal of the Japanese Society for Horticultural Science,61, 445–451.


Martinsen, P., & Schaare, P. (1998). Measuring soluble solid distribution inkiwifruit using near-infrared imaging spectroscopy. Postharvest Biology andTechnology, 14, 271–281.

Morimoto, S., McClure, W. F., & Stanfield, D. L. (2001). Hand-held NIR spec-trometry. Part I: An instrument based upon gap-second derivative theory.Applied Spectroscopy, 55, 182–189.

Morita, K., Shiga, T., & Taharazako, S. (1992). Evaluation of change in quality ofripening bananas using light reflectance technique. Memoirs of Faculty ofAgriculture, Kagoshima University, 28, 125–134.

Nussberger, S., Dekker, J. P., Kuhlbrandt, W., van Bolhuis, B. M., vanGrondelle, R., & van Amerongen, H. (1994). Spectroscopic characterization ofthree different monomeric forms of the main chlorophyll a/b binding proteinfrom chloroplast membranes. Biochemistry, 33, 14775–14783.

Osborne, B. G., & Feam, T. (1986). Near infrared data handling and calibration bymultiple linear regression. In Near infrared spectroscopy in food analysis.Harlow, UK: Longman Scientific & Technical.

Robert, P., Bertrand, D., Devaux, M. F., & Sire, A. (1992). Identification ofchemical constituents by multivariate near infrared spectral imaging.Analytical Chemistry, 24, 664–667.

Robert, P., Devaux, M. F., & Bertrand, D. (1991). Near infrared video imageanalysis. Sciences des Aliments, 11, 565–574.

SBIG. (1998). Product Catalog. Santa Babara, CA: SBIG Astronomical Instruments.

Sugiyama, J. (1999). Visualization of sugar content in the flesh of a melon by near-infrared imaging. Journal of Agricultural and Food Chemistry, 47, 2715–2718.

Taylor, S. K., & McClure, W. F. (1989). NIR imaging spectroscopy: measuring thedistribution of chemical components. In M. Iwamoto, & S. Kawano (Eds.),Proceedings of the 2nd International NIRS Conference (pp. 393–404). TokyoJapan: Korin.

Temma, M., Hanamatsu, K., & Shinoki, F. (1999). Development of a compactnear-infrared apple-sugar-measuring instrument and applications. InH. Tsuyuki (Ed.), Proceedings of the 15th Symposium on Non-destructiveMeasurements (pp. 113–117). Ibaraki, Japan: Japanese Society for FoodScience and Technology.

Tsuta, M., Sugiyama, J., & Sagara, Y. (2002). Near-infrared imaging spectroscopybased on sugar absorption band for melons. Journal of Agricultural and FoodChemistry, 50, 48–52.

Watada, A. E., Norris, K. H., Worthington, J. T., & Massie, D. R. (1976). Esti-mation of chlorophyll and carotenoid contents of whole tomato by lightabsorbance technique. Journal of Food Science, 41, 329–332.

hyperspectral imaging for food quality analysis and control || visualization of sugar distribution...

Documents