Articles
Tomato (Solanum lycopersicum) shape classification with deep learning AI-algorithms
Article number
1396_25
Pages
185 – 192
Language
English
Abstract
On the Belgian fresh market different cultivars of tomatoes (Solanum lycopersicum) are commercialized within different market segments according to size and shape.
An important tomato shape descriptor is called groovedness, a specific descriptor used to describe the amount and depth of the grooves or notches mainly found on the crown side of the tomato.
A 4-level scale was created for classification of this shape attribute: 1) tomatoes are perfectly round, 2) tomatoes appear lightly grooved, which indicates that some grooves are present, 3) tomatoes appear grooved and have more and deeper grooves that run through the equatorial plane, and 4) tomatoes are ribbed in which they have distinct notches in the fruit surface where the tomato has lost its round shape.
A data set of images of tomatoes that were scored for groovedness by an expert panel was collected to train a deep learning AI-algorithm.
A preprocessing step was applied to size and correct RGB-colour information before training the model.
Image augmentation consisted of rotating the images, which created additional input for the model.
Models using images taken from the crown side performed better then models using images from the equatorial plane.
A model accuracy of 75.3% was found using the crown side images, although 95% of the misclassified images only differed one groovedness class.
In general the lightly grooved tomatoes were more difficult to distinguish from the grooved tomatoes, while the extreme classes were more easily distinguished by the algorithm.
This also became clear when comparing the classifications performed by an expert panel as they too find it harder to distinguish between the middle classes.
Comparing the AI-algorithm with the expert panel showed that the algorithm was equally good at classifying images into the different classes, with the added benefit that the AI-algorithm is user independent.
An important tomato shape descriptor is called groovedness, a specific descriptor used to describe the amount and depth of the grooves or notches mainly found on the crown side of the tomato.
A 4-level scale was created for classification of this shape attribute: 1) tomatoes are perfectly round, 2) tomatoes appear lightly grooved, which indicates that some grooves are present, 3) tomatoes appear grooved and have more and deeper grooves that run through the equatorial plane, and 4) tomatoes are ribbed in which they have distinct notches in the fruit surface where the tomato has lost its round shape.
A data set of images of tomatoes that were scored for groovedness by an expert panel was collected to train a deep learning AI-algorithm.
A preprocessing step was applied to size and correct RGB-colour information before training the model.
Image augmentation consisted of rotating the images, which created additional input for the model.
Models using images taken from the crown side performed better then models using images from the equatorial plane.
A model accuracy of 75.3% was found using the crown side images, although 95% of the misclassified images only differed one groovedness class.
In general the lightly grooved tomatoes were more difficult to distinguish from the grooved tomatoes, while the extreme classes were more easily distinguished by the algorithm.
This also became clear when comparing the classifications performed by an expert panel as they too find it harder to distinguish between the middle classes.
Comparing the AI-algorithm with the expert panel showed that the algorithm was equally good at classifying images into the different classes, with the added benefit that the AI-algorithm is user independent.
Authors
D.J. Vanhees, M. Vanderbeken, B.E. Verlinden, B. Nicolaï
Keywords
artificial intelligence (AI), deep learning, tomato, classification, shape, image processing
Online Articles (82)