| Sumario: | Fungal infections are a main concern in fruit packing houses since one infected orange
can cause a whole storage room to be infected in case the infection is not detected. For
that reason, early detection before the infection is visible by the human eye is crucial. In
many packing houses, UV light is used to manually sort oranges, but it can be dangerous
for the eyes. Besides, at the very beginning of the infection, it can be difficult to detect it
even using UV light, and since the human eye gets exhausted when working for hours,
that makes human early detection not suitable for the task. However, hyperspectral imaging
can solve that problem since it avoids the use of UV light and humans due to its
ability to detect not only external but also internal information of the fruit.
For that reason, a preliminary study was carried on using 100 oranges of Navel-late
variety. 50 of them were inoculated with distilled water as a control, and 50 of them with
Penicillium Digitatum with a concentration of 106 spores ml-1. After that, they were
stored at 20ºC and 90% relative humidity for 4 days. Images were taken at the day of
inoculation and every day for 4 days after inoculation, taking a total of 500 images.
Hyperspectral images of these fruits were acquired with a VIS/NIR pushbroom hyperspectral
imaging system (900-1700 nm) and were corrected with the white and dark
reference and the background was removed. Images were then divided into train (60%
of the images), validation (20% of images) and test (20% of images) datasets of healthy
and inoculated images.
Unlike most authors who use 2D CNNs, we have used a 3D customized CNN which is
designed for 3-dimension data, and which allows obtaining more spectral information
than 2D CNNs. Images were resized to 100 x 100 pixels due to computational limitations
and data augmentation was applied before training to increase the number of samples.
Results showed that after one day, the model was able to classify correctly 100% of the
control oranges and 80% of the inoculated ones, proving that this technology might be
able to predict infection after it is visible by the human eye and even when using UV light.
However, this should and will be studied in a future using more samples and better computational
resources as to improve the model and the image resolution.
|
|---|