Low requirement imaging enables sensitive and robust rice adulteration quantification via transfer learning. (September 2021)

• Record Type: Journal Article
• Authors: Pradana-López, Sandra; Pérez-Calabuig, Ana M.; Rodrigo, Carlos; Lozano, Miguel A.; Cancilla, John C.; Torrecilla, José S.
• Is Part Of: Food control. Volume 127(2021)
• Publication Date: 2021-09
• Subjects: Rice adulteration quantification -- Low requirement imaging -- ResNet34; Food -- Quality -- Periodicals; Food -- Analysis -- Periodicals; Food handling -- Periodicals; Food industry and trade -- Quality control -- Periodicals; Aliments -- Industrie et commerce -- Qualité -- Contrôle -- Périodiques; Aliments -- Qualité -- Périodiques; Aliments -- Analyse -- Périodiques; Hygiène alimentaire -- Périodiques; Food -- Analysis; Food handling; Food -- Quality; Periodicals; Electronic journals; 664.07
• DOI: 10.1016/j.foodcont.2021.108122
• Languages: English
• ISSNs: 0956-7135

Abstract

Abstract: In order to develop a rice adulteration detection system, a deep learning method was implemented to classify simple photographs of five different types of rice. Firstly, the different types of rice were milled and sieved, enabling the imaging of not only grain, but also rice in flour format. Pure rice types as well as mixtures in different percentages (25%, 50%, and 75%) were photographed to build the database. A basic camera was used to capture different images of the samples reaching a total of 3400 photos. As far as the mathematical algorithm is concerned, a transfer learning based ResNet34 was employed to classify the rice into their unique groups. Using a randomly selected 90% of the total database for training and internal validation, an overall accuracy of 98.0% was obtained after averaging the individual performance for each of the 34 analyzed classes. Finally, a blind test was performed with the remaining 10% of the images, reaching a 98.8% correct classification rate. Highlights: Grain and flour images of rice captured with a simple camera. Transfer learning implemented to effectively fight adulteration and food fraud. Up to 34 classes of pure and adulterated rice identified accurately. Blinded samples classified correctly at a 98.8% rate. Quality control of rice enabled for real-time and inexpensive analysis.