Development and application of a model for the automatic evaluation and classification of onions (Allium cepa L.) using a Deep Neural Network (DNN)

Abstract

Evaluating onions for size, shape, damage, colour and discolouration is the first and most important step in classifying them for raw material quality, processing and the horticultural and agri-food sectors. Current methods of geometric evaluation and grading of onions involve mechanical and extremely invasive sorting, which causes additional damage, reduces the quality of the raw material and is also labour and time-consuming. As a result, non-invasive evaluation and classification methods that are both fast and accurate are being sought. One such method is digital image analysis, which, when combined with instrumentation and deep neural networks, can fully automate the process. The main aim of this study was the development of a model for the automatic evaluation and classification of onions using a deep convolutional neural network (CNN) model. A fixed-architecture network was built, for which a computational algorithm was developed in Python 3.9 and published at https://github.com/piotrrybacki/onion-CNN.git (accessed on 4 October 2024). The Hyduro F1 onion variety, a hybrid all-purpose variety of the Rijnsburger type, was used to build, teach and test the model. The developed algorithm classified the onion images qualitatively with an accuracy of 91.85%. This classification was based on the geometric parameters of the onion, i.e. diameter, height, transversal and longitudinal circumference, and the estimated area of damage or discolouration of the skin. The root mean square error (MSE) in RGB space varied between 87.99 and 91.24, and the maximum image classification time was 28.98 ms/image. The developed algorithm has a very high utility, as it automates the classification process, reducing its time and labour intensity.