Selection of Optimal Wavelength Features for Decay Detection in Citrus Fruit Using the ROC Curve and Neural Networks

• Published in: Food and bioprocess technology Vol. 6; no. 2; pp. 530 - 541
• Main Authors: Lorente, Delia, Aleixos, Nuria, Gómez-Sanchis, Juan, Cubero, Sergio, Blasco, Jose
• Format: Journal Article
• Language: English
• Published: New York Springer-Verlag 01-02-2013; Springer Nature B.V
• Subjects: Agriculture; Algorithms; Artificial neural networks; Biotechnology; Chemistry; Chemistry and Materials Science; Chemistry/Food Science; Citrus; Citrus fruits; Classification; Decay; Economic impact; Extreme values; Food Science; Fruits; Hyperspectral imaging; Image classification; Image detection; Learning algorithms; Lighting; Machine learning; Mandarins; Multilayer perceptrons; Neural networks; Original Paper; Scars; Citrus fruits; Computer vision; Non-destructive inspection; ROC curve; Decay; Hyperspectral imaging
• ISSN: 1935-5130; 1935-5149
• DOI: 10.1007/s11947-011-0737-x

Early automatic detection of fungal infections in post-harvest citrus fruits is especially important for the citrus industry because only a few infected fruits can spread the infection to a whole batch during operations such as storage or exportation, thus causing great economic losses. Nowadays, this detection is carried out manually by trained workers illuminating the fruit with dangerous ultraviolet lighting. The use of hyperspectral imaging systems makes it possible to advance in the development of systems capable of carrying out this detection process automatically. However, these systems present the disadvantage of generating a huge amount of data, which must be selected in order to achieve a result that is useful to the sector. This work proposes a methodology to select features in multi-class classification problems using the receiver operating characteristic curve, in order to detect rottenness in citrus fruits by means of hyperspectral images. The classifier used is a multilayer perceptron, trained with a new learning algorithm called extreme learning machine. The results are obtained using images of mandarins with the pixels labelled in five different classes: two kinds of sound skin, two kinds of decay and scars. This method yields a reduced set of features and a classification success rate of around 89%.