| Sumario: | Fungal infections are a main concern in fruit packing houses since a single infected fruit can
spread the infection, causing severe losses. Hence, early detection is crucial. Fluorescence
induced by UV light is commonly used to detect infected fruits, but it can harm the eyes and
the skin. Hyperspectral imaging and chemometrics have already been used to detect early
infected fruits, but novel deep-learning approaches can improve the detection. In this work, 50 oranges Navelate were inoculated with Penicillium digitatum and another 50 oranges were
inoculated with distilled water as a control. After the inoculation, the fruits were stored at 20
ºC and 90 % relative humidity. Images were taken daily from the inoculation for four days,
using a Vis-NIR pushbroom system (900-1700 nm). A total of 500 images were captured,
corrected using white and dark references, cropped to the size of the fruit and divided into a
train (60% of the images), validation (20% of images) and independent test (20% of images)
datasets. A customised convolutional nerual network (CNN) was designed, composed of 3
convolutional layers (16, 32 and 64 filters and a kernel size of 3), followed by 3 max-pooling
layers (pool size 2) and two final dense layers for classification of 128 and 1 neuron (with
sigmoid activation function) which was trained for 800 epochs with a binary cross-entropy loss
function. It was possible to correctly classify 100% of the control oranges and 80% of the
inoculated ones from the second day, demonstrating that it is possible to detect asymptomatic
infected using this technology.
|
|---|