Application of digital imaging analysis and pattern recognition in edible oils classification by using color changes during primary and secondary oxidation process

Discerning the expiration status (rancid and non-rancid) of edible vegetable oils is very significant because of hazardous primary and secondary oxidation products. Oils are a nutritious and valuable food source which play an important role not only in supplying energy but also in sustaining a health. Edible vegetable oils such as soya, sunflower, canola, sesame and olive, bring essential nutrient components for human being such as vitamins, fatty acids, and micronutrients, which are necessary for daily life. Lipid oxidation in vegetable oils is associated with unsaturation of the oils. This reaction leads to the formation of a series of intermediate compounds named hydroperoxides. Hydroperoxides are the primary oxidation products of lipid oxidation. Due to the unstable nature of these primary products which leads to their decomposition and turning into secondary oxidation products, such as carbonyl compounds occur soon.
The use of expired edible oils leads to a decrease in the nutrition value and an increase in potential hazards to people's health, so monitoring the quality and security of edible oils is important. Based on the reports and experimental observation the oil color changed during oxidation. Therefore, it is of utmost importance to find new and fast methods for detecting the quality of oils. Computer vision in food sciences is an affordable technology and is extensively used. The aim of this study was to introduce a simple and feasible method for classifying edible vegetable oils (soya, sunflower, canola, sesame and olive) and also for distinguishing their quality in terms of rancidity. In order to achieve this, multivariate statistical methods based on their rejection point of primary and secondary oxidation products was implemented.
Digital camera and unsupervised multivariate statistical techniques such as linear discriminant analysis (LDA) and principal component analysis (PCA) were used for pattern recognition and classification. In this study, the physicochemical characterization of 77 oil samples includes their peroxide and carbonyl values were evaluated at 80 ◦C. The color indices L*a*b* were used for this classification. The space that was built for imaging was 120cm ×90cm ×90cm with dark walls to isolate the samples from external light. The compartment has a camera (Canon model, EOS 1000D), which was connected to computer by USB port. The illumination of the compartment was performed by using eight fluorescent lamps with 8 W (white color), the lamps were placed at a distance of 20 cm from the samples. The illustration was performed by Zoombrower EX 0.5, the other characteristics of the camera for imaging were as follow: flash (off), zoom (on), Iso speed (100), Aperture priority (F / 20) and Shutter speed (0.6 Sec). The illumination condition at compartment for each sample was the same. Image color analysis was performed using the Image j (Version: 1.4.3.67) software to convert images from R*G*B color space to L*a*b. The recorded images contained 24-bit (16.7 million colors) and 3888 pixels × 2592 pixels spatial resolution and were stored in JPEG format (jpg). A specific region at the center of each image was selected for converting R*G*B to L*a*b. In this study, three components of color space L*a*b* were extracted from 231 images samples ( 77 images of different types of oil before heating, 77 images at the rejection point based on peroxide value and 77 images on the rejection point based on carbonyl value). The extracted color values were used for linear discriminant analysis classification and principal component analysis. The classification was performed using MATLAB (R2013) software
The comparison of the results of the linear discriminant analysis showed that distinguishability between the two types of different oils was 100% and only the distinguishing of one oil type in rancid and non-rancid state resulted in a decrease in accuracy to 97%. Also the overall and simultaneous analysis of oil samples in both states (rancid and non-rancid) by the two classifiers of LDA and PCA showed that the classification of each oil individually has the highest accuracy (100%) and the results of the studying several different oils showed a decreased accuracy (98% and 96%). However, in practice, the result of this classification given the diverse colour range of vegetable oils, is acceptable in terms of accuracy and the linear discriminant analysis classifier acted more successfully compared to principal component analysis classifier by about 40%.