Ensayo sobre el futuro

Statistical pattern recognition classification with computer vision images for assessing the furan content of fried dough pieces

Gabriel A. Leiva-Valenzuela1, *, María Mariotti2, Germán Mondragón3, Franco Pedreschi1

1 Department of Chemical and Bioprocess Engineering, Pontificia Universidad Católica de Chile, Avenida Vicuña Mackenna 4860, Macul, Santiago, Chile.

2 Department of Biotechnology, Universidad Tecnológica Metropolitana, Avenida José Pedro Alessandri 1242, Ñuñoa, Santiago, Chile

3 Informatic and Machine Vision Department, Quality For Technology Limitada, Guardia Vieja 181 oficina 506, Providencia, Santiago, Chile.

* Corresponding author, phone: (+562) 2354-4264, e-mail: gmleiva@uc.cl

Abstract

Furan is a toxic derived from the non-enzymatic browning reaction (NEBR). Scientists are concerned about limiting dietary exposure of furan; however, the expensive and time-consuming analytical techniques for furan are unrealistic to implement for in-line applications. The objective of this paper was to test furan classification models in fried starchy matrices based on the application of pattern recognition of images. Starchy doughs were fried at 150, 160, 170, 180, and 190°C for 5, 7, 9, 11, 13, and 30 min; the furan content was measured, and some color images were extracted. To train the classifier, the furan content was quantified by gas chromatography–mass spectrometry (GC–MS). Corresponding images were acquired and processed to extract 2175 chromatic and textural features. Principal component analysis (PCA) was used to reduce the dimensionality to 8–12 principal components. In parallel, sequential forward selection (SFS) coupled with linear discriminant analysis (LDA) was the best strategy to select only 5-7 features, being mean intensity measured in red channel (R) and local binary patterns (LBP) the most important. Remarkable results show that LDA resulting in the best classifier, 91.39-97.60 % of samples above 113 µg/kg of furan level under 10-fold cross validation and 69.54-83.80 % of class 1(0-38 µg/kg) from class 2 (39-113 µg/kg). Finally, support vector machine (SVM) recognized 87.71-96.74% of class 3 (114-398 µg/kg) from class 4 (399-646 µg/kg). The computer vision may be used to detect high amount of furan in fried starchy matrices.

Keywords: Maillard reaction, starchy foods, image processing, non-enzymatic browning.

1. Introduction

Cooking, baking, toasting, roasting, and frying are common thermal treatments used to make tasty and nutritive foods. These treatments involve chemical changes that are promoted by reactions, such as protein denaturation, sugar caramelization, and non-enzymatic browning (NEB), that are crucial for the development of positive sensorial attributes, such as flavor, color, and texture (Anese and Suman, 2013; Wang et al., 2013; Martins et al., 2000; Tareke et al., 2002). Unfortunately, NEB also promotes the formation of neo-contaminants, such acrylamide, 5-hydroxymethylfurfural, and furan, negatively affecting the chemical safety of foods.

Furan (C4H4O) is a small lipophilic organic compound (MW = 68 g mol−1) (Hasnip et al., 2006) and a possible human carcinogen that is found in several foods processed at high temperatures, such as coffee and baby foods (Crews and Castle 2007, Mariotti et al., 2012). For this reason, efforts have focused on understanding the mechanism of furan formation. Thermal degradation and the rearrangement of sugars were suggested as the primary sources of furan in food (Fan, 2005; Morehouse et al., 2008). However, amino acids, polyunsaturated fatty acids, and ascorbic acid have also been implicated as critical precursors (Maga, 1979; Van Lancker et al., 2009; Vranova and Ciesarova, 2009; Zoller et al., 2007).

Since there is no intake limit of furan, dietary exposure to furan from common foods has been previously reported. The Food and Drugs Administration of the United States of America (FDA) has estimated the average intake for United States consumers to be 0.2 µg/kg bw/d, detecting that furan in slow moisture foods ranged from < 0.2 parts per billion (ppb) to over 170 ppb. Complementary antecedents regarding dietary exposure to furan showed the high impact of starchy foods, as presented by Mariotti et al. (2012), who reported furan exposure for different age groups in Chile, where school children (10–13 years old) were exposed to the highest levels of furan (~500 ng kgbw−1 day−1) compared to nine-month-old babies (~250 ng kgbw−1 day−1), adults (~70 ng kgbw−1 day−1), and the elderly (~100 ng kgbw−1 day−1). One of the causes of high furan exposure of school children was the high intake of frying starchy foods in their regular diets.

Sadly, misinformation, a nonexistent regulatory framework, and the high costs of the regular industrial estimation of furan content can impact incipient furan assessment in the food industry. The traditional method of determining the furan content, which is based on gas chromatography–mass spectrometry, is quite expensive and time-consuming. Moreover, equipment cost, specific reagents, and technical knowledge restrict this methodology to research centers and analytical laboratories, as furan testing is unaffordable in industry. It is paramount to develop non-destructive methods that are easy-to-implement in an industrial setting.

Computer vision (CV), has been applied to food quality evaluation in the last ten years; however, studies that use CV have concentrated on using or developing tailored methods based on visual features that are able to complete a specific task, such as to identify color or shapes (Mery et al., 2013). Accordingly, CV has been one of the most useful methodologies to evaluate the external quality of agriculture products and processing foods. The analysis of an image of a food product by computers allows for the control of processes and study of phenomena in foods (Leiva-Valenzuela and Aguilera, 2013).

Although CV techniques have been used many times in starchy food quality operations (Matiacevich et al., 2012; Pedreschi et al., 2012), there have been few applications in food safety systems to identify and quantify toxic compounds in easy, fast, and connected ways. Pedreschi et al. (2006) implemented a CV system to evaluate the color of fried potato slices using color images to describe the NEB. CV has also been used to study the kinetics of acrylamide formation and its correlation with surface color in fried potatoes (Pedreschi et al., 2007a; Pedreschi et al., 2007b; Pedreschi et al., 2007c; Pedreschi et al., 2005). More recently, good correlations were found between the acrylamide content of fried potatoes and their color (Serpen and Gökmmen, 2009). These studies reinforce the idea that the color of foods, specifically the surface color of potato chips, is highly correlated with the acrylamide content. Interestingly, some authors have acquired potato chip reflectance images by near-infrared spectroscopy (NIR) to simultaneously estimate the content of fat, dry matter, and acrylamide in fried potato chips (Pedreschi et al., 2010). Although the prediction error was fairly high, this study demonstrates the potential of using visible and near-infrared spectroscopy as screening techniques for evaluating the acrylamide content in potato chips. Since acrylamide was the first new contaminant from NEB described since 2002 in potato chips, most studies have focused on this toxic food compound.

...