See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/286916474

Detection of Fruit-fly Infestation in Olives using X-Ray Imaging: Algorithm

Development and Prospects

Article · January 2016

DOI: 10.7726/ajast.2016.1001

CITATIONS READS

2 542

4 authors, including:

R.P. Haff Roberto Moscetti

United States Department of Agriculture Tuscia University
73 PUBLICATIONS   961 CITATIONS    61 PUBLICATIONS   545 CITATIONS   

SEE PROFILE SEE PROFILE

Riccardo Massantini
Tuscia University
101 PUBLICATIONS   1,425 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

SANCAST - Quality and health of the chestnut View project

Persimmon View project

All content following this page was uploaded by Roberto Moscetti on 02 June 2016.

The user has requested enhancement of the downloaded file.

Columbia International Publishing
American Journal of Agricultural Science and Technology
(2015) Vol. 4 No. 1 pp. 1-8
doi:10.7726/ajast.2016.1001
Research Article

Detection of Fruit-fly Infestation in Olives using X-Ray

Imaging: Algorithm Development and Prospects

Ronald P. Haff1*, Eric S. Jackson1, Roberto Moscetti2, and Riccardo Massantini3

Received 2 July 2015; Published online 12 December 2015

© The author(s) 2015. Published with open access at www.uscip.us

Abstract
An algorithm using a Bayesian classifier was developed to automatically detect olive fruit fly infestations in x-
ray images of olives. The data set consisted of 249 olives with various degrees of infestation and 161 non-
infested olives. Each olive was x-rayed on film and digital images were acquired with a film scanner at a
resolution of 59 pixels per cm. Features extracted from the images were submitted to the classification
algorithm and error rates for detection of the infestations obtained. Feature selection involved pixel intensity
values and pixel derivative values at each pixel location in the image. The ability of the algorithm to
differentiate infested and non-infested olives was tested. Internal damage to the olive was a factor in
detection, with slight damage correctly identified 50% of the time and severe damage correctly identified
86% of the time. Non-infested olives were correctly identified with 90% accuracy.

Keywords: Olives; Fruit-fly; X-ray Imaging; Bayesian classifier; Feature selection

1. Introduction
Olives and olive oil are important global commodities. Roughly three million tons of oil alone was
produced in 2012, of which approximately 90% was in Spain, Italy, or Greece (IOC, 2013). In the
United States, virtually all olives are grown in California, which on a global scale is a small but
rapidly growing producer. California production grew from 46,300 tons in 2009 to 166,000 tons in
2013, peaking in 2012 at 180,000 tons (USDA-NASS, 2014). This represents a roughly 400%
increase over a three-year period. California olives are split approximately evenly between table
olives and olive oil.

The chemical composition of virgin olive oil differs from other vegetable oils in a number of
______________________________________________________________________________________________________________________________
*Corresponding e-mail: [email protected]
1* United States Department of Agriculture, Agricultural Research Service, Western Regional Research
Center, Albany, CA.
2 Department of Science and Technology for Agriculture, Forest, Nature and Energy, Tuscia University,
Viterbo, Italy
3 Department for Innovation in Biological, Agro-food and Forest system, Tuscia University, Viterbo, Italy

33
 Ronald P. Haff, Eric S. Jackson, Roberto Moscetti, and Riccardo Massantini / American Journal of Agricultural
Science and Technology (2015) Vol. 4 No. 1 pp. 1-8

significant ways that affect sensory, nutritional and health properties. Studies have suggested that
extra-virgin olive oil, a key component of the Mediterranean diet, is a factor in reduced rates of
heart disease in the region (Katan et al., 1995). Since olive fruit quality is directly correlated with
the chemical composition and overall quality of produced oil, methods for removing defective fruit
would benefit producers and consumers alike.

Bactrocera oleae (olive fruit fly) is the main insect pest of olives worldwide (Daane and Johnson,
2010) and a frequent cause of reduced olive oil quality. The first discovery of the olive fruit fly in
California occurred in 1998 and it now poses a significant threat to commercial olive growers
throughout the state. Significant state and federal resources have been devoted to controlling the
spread of the olive fruit fly in California, reflecting the significance of the problem. This includes
funding allocated by Congress, local marketing associations such as the California Olive Commission
and the California Department of Food and Agriculture (Olive Oil Source, 2006). The adult female
lays eggs near the surface of the fruit and the developing larvae feeds on the fruit, excavating
tunnels that can reach the seed (Rice, 2000). The injured fruit is then susceptible to subsequent
bacterial and fungal infestation, resulting in loss of fruit integrity and oil quality. Infestation by the
olive fruit fly reduces oil yield and negatively impacts free acidity, peroxide value and ultraviolet
absorption (Gómez-Caravaca et al., 2008). Reduced antioxidant phenolic content (Gómez-Caravaca
et al., 2008; Gucci et al., 2012) and altered volatile compound profiles (Angerosa et al., 2004)
leading to off-flavors have also been observed. Consequently, the nutritional value and the sensory
properties of oil extracted from infested olives are compromised, and the product often does not
conform to legal specification for extra-virgin or virgin oils.

California table fruit processors have zero tolerance for olive fruit fly damage (Vossen et al., 2005;
Yokoyama, 2012), and the presence of even a few infested fruits can lead to rejection of an entire lot
(Zalom et al., 2003). Since not all olives in an infested lot are infested, an excellent product could be
recovered if good product could be separate from defective product. Traditionally, research on the
olive fruit fly has focused on control strategies including monitoring, trapping, sanitation, bait
sprays, and biological controls (Zalom et al., 2003). Until recently, no research was reported on
detection and removal of insect damaged fruit from the processing stream. This has changed;
however, with Moscetti et al. (2015) demonstrating the feasibility of detecting infested olives using
NIR spectroscopy to determine a small number of wavebands with discriminating power.

Figure 1 shows photographs and x-ray images of infested and non-infested fruit for comparison of
surface markings and the corresponding damage inside the olive. Generally, infested fruit has a
small hole on the surface at the point of exit, suggesting machine vision as a potential means of
detection. However, preliminary attempts have shown this to be unreliable because other kinds of
surface damage leave similar marks that are difficult to distinguish from insect damage. While the
markings on the surface are similar, the interiors are not. The olive with the exit hole shows
extensive damage due to tunneling, while the non-infested olive shows negligible interior damage.
Thus, surface imaging is not a reliable technique for detection of infestations.

34
 Ronald P. Haff, Eric S. Jackson, Roberto Moscetti, and Riccardo Massantini / American Journal of Agricultural
Science and Technology (2015) Vol. 4 No. 1 pp. 1-8

Fig. 1. Picture and film x-ray images of olives showing a) an olive fruit fly exit hole on the surface
and extensive damage from olive fruit fly tunneling, compared with b) surface markings
similar to an exit hole but containing negligible damage inside the olive.

X-ray imaging is ideally suited for detection of interior defects and contaminants in agricultural
commodities. Inspection of x-ray film over a light table is an available method to estimate the
amount of hidden infestation present. However, sample sizes must be kept small due to the
intensive labor involved and the expense of film and developing. Furthermore, results could be
inconsistent due to the differences between inspectors and factors such as fatigue. Automated
detection algorithms are thus an essential component in imaging methods. Discriminant analysis
using Bayesian classifiers is well suited for high-speed image analysis, depending on the number of
features that are being selected and the amount of computation required deriving those features.
One challenge is to identify features that can be extracted with minimal computation. Another is to
restrict the number of features to as small a set as possible while still retaining the essential
information that allows proper separation into classes.

Irradiation of food products is associated with legitimate concerns about effects on quality,
including the development of free radicals and nutrient loss (IFT, 2015). However, it is important to
differentiate between food irradiation, which generally involves doses on the order of KGy, and x-
ray imaging as reported here, which involves doses on the order of mGy. Thus, there is no
measurable effect on food quality or safety as a consequence of x-ray imaging.

The objective of this study was to evaluate the potential use of x-ray imaging to detect olive fruit fly
infestation and to develop an algorithm based on Bayesian Classification to automatically detect
infestations in x-ray images.

35
 Ronald P. Haff, Eric S. Jackson, Roberto Moscetti, and Riccardo Massantini / American Journal of Agricultural
Science and Technology (2015) Vol. 4 No. 1 pp. 1-8

2. Methods and Materials

2.1 X-ray Imaging

Infested and non-infested olives were collected and imaged on x-ray film using an x-ray cabinet
(43804N, Faxitron X-ray Corporation, Wheeling, IL) at 28 keV, 3mA for 180 seconds. The resulting
film was visually inspected to determine which olives were infested. The damage due to infestation
was subjectively classified as slight, moderate, or severe. Figure 2 shows sample x-ray images for
each classification.

Fig. 2. X-ray images of olives divided into four levels of fruit fly damage.

The data set consisted of 410 images, 249 of which were infested and 161 non-infested. A digital
image of each olive was generated using a film scanner (ScanMaker 9800XL, Microtek, Carsen, CA)
at a resolution of 59 pixels per cm. The set of images was split into two, one half for algorithm
training and the other for validation. The distribution of images between the training and validation
sets, and among infestation damage classifications, are shown in Table 1. The uneven distribution of
samples is a consequence of the limited number of samples available.
Table 1 Data set distribution.
Infestation damage Training set Validation set Total
Slight 15 16 31
Moderate 23 23 46
Severe 86 86 172
Non-infested 81 80 161
Total 205 205 410

36
 Ronald P. Haff, Eric S. Jackson, Roberto Moscetti, and Riccardo Massantini / American Journal of Agricultural
Science and Technology (2015) Vol. 4 No. 1 pp. 1-8

2.2 Bayesian classifiers

Sixty-four features were selected from each olive image as values within the bins of a two
dimensional histogram. The histogram bins contained the number of image pixels that fell within
discreet levels of derivative and intensity values (Fig. 3). This method of feature selection has been
applied successfully to distinguish defects in pistachio nuts (Pearson et al., 2001), as well as
infested from non-infested wheat kernels (Haff, 2001). These are similar problems to separating
infested from non-infested olives.

Fig. 3. Method of creating a two-dimensional feature histogram. Each pixel in the image (neglecting
boundary considerations) contributes a single count to one of the 64 bins in the histogram,
the final values of which comprise the features submitted to the classifiers. Note that in the
above image, blank pixels have zero intensity (background) and for clarity the pixel
intensities (and hence the derivative values) have been limited to a small set of values. In
this example the derivative step size is four.

The derivative at a specific pixel location was computed as the absolute difference between the
pixel intensities a specified distance (in pixels) in the forward and backward horizontal direction
(Fig. 3). The optimal step distance was determined by trial and error to be three pixels for this data
set. No smoothing was implemented in obtaining the derivative values. The pixel intensity ranges
and derivative values defining the histogram bins were also selected by trial and error based on
error rates obtained when the resulting features were submitted to a discriminant analysis routine.
While it is ideal to use all 64 features for best results, for real-time applications where high
throughput is desired this is not always computationally feasible. A computer algorithm was
therefore developed to test all combinations of three features for best recognition results using
discriminant analysis. The training set of images was used to compute the discriminant function,
which was then applied to the validation set to obtain recognition error results.

The algorithm was tested for two separate conditions. First, all infested images were used for both
training and validation. It was anticipated that the small number of images for the slight and

37
 Ronald P. Haff, Eric S. Jackson, Roberto Moscetti, and Riccardo Massantini / American Journal of Agricultural
Science and Technology (2015) Vol. 4 No. 1 pp. 1-8

moderate classifications might lead to poor results, so a second test was done using only those
olives with severe damage.

3. Results and Discussion

The results for testing the algorithm with all levels of infestation damage are shown in Table 2. As
expected, the recognition was low for slight and moderate damage, at 50% and 52% respectively.
For severe damage the recognition rose to 87%. False positives (good product classified as bad)
were 16%. Total error rate for infested images was 24%. As stated earlier, the poor results and high
false positives are believed to be a consequence of the small number of slight and moderate
classifications in the training set, hindering the effort to derive a proper decision boundary between
the two regions.

Table 2 Recognition results using images for all stages of infestation.

Infestation Damage # in validation set # correctly classified % Recognition
Slight 16 8 50
Moderate 23 12 52
Severe 86 75 87
Non-infested 80 67 84
Total 205 162 79

The results obtained testing the algorithm using only the severely damaged olives are shown in
Table 3. As expected, the recognition of severe damage remained almost unchanged, while false
positives dropped from 16% in the previous case to 10%. This is because the absence of the slight
and moderate damage in the training process allowed a more accurate derivation of the decision
boundary. We are predominantly interested in the severe damage caused by infestation, although
detection of all damage would be ideal. However, when balancing the desire to detect all damage
with the need to reduce the false positive rate, the latter is generally considered more important as
high false positive rates lead to the rejection of substantial portions of the crop. Future work will
concentrate on reducing the false positive rate, which presumably can be accomplished by training
with a larger number of samples so that a more precise decision boundary can be achieved.

Table 3 Recognition results using only images of severe infestations.

Infestation damage # in validation set # correctly classified % Recognition
Severe 86 74 86
Non-infested 80 72 90
Total 166 146 88

One of the benefits of the feature selection process used here is that it requires no prior knowledge
of what features in the image are significant, as the process will sort them out. The technique
should work just as well for any sets of images that can be sorted into distinct classes. There may be
more suitable criteria for setting up the feature histogram than derivatives and pixel intensities,
and the spacing of the histogram bins is an important factor that must be considered.

38
 Ronald P. Haff, Eric S. Jackson, Roberto Moscetti, and Riccardo Massantini / American Journal of Agricultural
Science and Technology (2015) Vol. 4 No. 1 pp. 1-8

This research used high resolution film based x-ray images, which would not be practical for high
speed sorting of olives as film is relatively slow in terms of exposure and developing time.
Alternatives to film exist, including linescan systems and CCD or phosphor based detector systems.
These systems can achieve the high imaging speeds required for on-line inspection, but generally
suffer in terms of image quality due to high noise as a consequence of extremely short exposure
times. As technology improves it is expected that these limitations will be overcome, and the
algorithm developed here will be applicable for real-time sorting systems.

4. Conclusion

The feasibility of using x-ray imaging to detect olive fruit fly infestations has been established.
Human observation of x-ray images can detect even the smallest hidden infestations. An algorithm
was developed to automatically detect infestations in x-ray images of olives. When detecting only
severe damage due to infestation, 86% of infestations were correctly classified while maintaining
false positives below 10%. Improvements in the algorithm are required to improve detection of
smaller amounts of damage and reduce the false positive rate. The principles used in this study
could be applied with real-time digital x-ray equipment as the basis for a sorting device.

References
Angerosa, F., Servili, M., Selvaggini, R., Taticchi, A., Esposto, S., & Montedoro, G. (2004). Volatile compounds
in virgin olive oil: occurrence and their relationship with the quality. Journal of Chromatography A,
1054, 17–31.
http://dx.doi.org/10.1016/S0021-9673(04)01298-1
Daane, K. M., & Johnson, M. W. (2010). Olive fruit fly: managing an ancient pest in modern times. Annual
Review of Entomology, 55, 151–169.
http://dx.doi.org/10.1146/annurev.ento.54.110807.090553
Gómez-Caravaca, A. M., Cerretani, L., Bendini, A., Segura-Carretero, A., Fernández-Gutiérrez, A., Del Carlo, M.,
Compagnone, D., & Cichelli, A. (2008). Effects of fly attack (Bactrocera oleae) on the phenolic profile and
selected chemical parameters of olive oil. Journal of Agricultural Food Chemistry, 56, 4577–4583.
http://dx.doi.org/10.1021/jf800118t
Gucci, R., Caruso, G., Canale, A., Loni, A., Raspi, A., Urbani, S., Taticchi, A., Esposto, S., & Servili, M., (2012).
Qualitative changes of olive oils obtained from fruits damaged by Bactrocera oleae (Rossi). HortScience
47, 301–306.
Haff, R. P. (2001). X-ray inspection of wheat for granary weevils. Ph.D. dissertation. Davis, CA: University of
California, Department of Biological and Agricultural Engineering.
Institute of Food Technologists (IFT). (2015). Irradiation of Food. http://www.ift.org/knowledge-
center/read-ift-publications/science-reports/scientific-status-summaries/irradiation-of-food.aspx.
International Olive Council (IOC). (2014). http://www.internationaloliveoil.org/estaticos/view/131-world-
olive-oil-figures.
Katan, M. B., Zock, P. L. & Mensink, R. P. (1995). Dietary oils, serum lipoproteins, and coronary heart disease.
American Journal of Clinical Nutrition, 61, 1368–1373.
Keagy, P.M., & Schatzki, T.F. (1993). Machine recognition of weevil damage in wheat radiograms. Cereal
Chemistry, 70(6), 696-700.
Moscetti, R., Haff, R. P., Stella, E., Contini, M., Monarca, D., Cecchini, M., & Massantini, R. (2015). Feasibility of

39
 Ronald P. Haff, Eric S. Jackson, Roberto Moscetti, and Riccardo Massantini / American Journal of Agricultural
Science and Technology (2015) Vol. 4 No. 1 pp. 1-8

NIR spectroscopy to detect olive fruit infested by Bactrocera oleae. Postharvest Biology and Technology,
99, 58-62.
http://dx.doi.org/10.1016/j.postharvbio.2014.07.015
Olive Oil Source. (2006). Olive Fly Control. http://www.oliveoilsource.com/page/olive-fly-control.
Pearson, T. C., Doster, M., & Michailides, T. J. (2001). Automated sorting of pistachio defects by machine
vision. Applied Engineering in Agriculture, 17(5), 729-732.
http://dx.doi.org/10.13031/2013.6905
Rice, R. E. (2000). Bionomics of the olive fruit fly, Bactrocera (Dacus) oleae. University of California
Cooperative Extension, UC Plant Protection Quarterly, 10, 1–5.
USDA-NASS. (2014). National Agricultural Statistics Database. http://quickstats.nass.usda.gov/.
Vossen, P., Varela, L., & Devarenne, A. (2005). Olive Fruit Fly. Univerisity of California Cooperative Extension
Sonoma County.
Yokoyama, V. Y. (2012). Olive fruit fly (Diptera: Tephritidae) in California: longevity, oviposition, and
development in canning olives in the laboratory and greenhouse. Journal of Economic Entomology,
105(1), 186-195.
http://dx.doi.org/10.1603/EC11255
Zalom, F. G., Van Steenwyk, R.A., & Burrack, H. J. (2003). Olive Fruit Fly. Pest Notes Publication 74112,
University of California Extension Service, Agricultural and Natural Resources.
http://fruitsandnuts.ucdavis.edu/files/74135.pdf.

40

View publication stats