Non-targeted approach to detect pistachio authenticity based on digital image processing and hybrid machine learning model

In this study, we propose a new method for detecting green pea adulteration in pistachio based on digital image and machine learning (ML). An algorithm was built using digital image processing techniques to detect region of interest (ROI) on adulterated pistachio images and a hybrid ML to classify the level of adulteration as class 1 (%0), class 2 (%10), class 3 (%20), class 4 (%30), class 5 (%40), and class 6 (%50) in a fully automated way. A dataset with size of 1254 × 15 were created. Training set and test set with the rate of 80% and 20% based on fivefold cross validation were created. Decision tree, random forest (RF), k-nearest neighboring, support vector machines, naïve bayes and artificial neural network (ANN) are performed and compared to classify the level of adulteration in two steps as direct and binary classification. ANN has achieved the highest results as 93.65% of accuracy and 0.87 of Matthews correlation coefficient (MCC) based on direct classification to separate class1, class 2, class 5, and class 6 from class 3 and class 4. RF has achieved the highest results as 89.56% of accuracy and 0.79 of MCC based on binary classification to separate class3 from class 4. As a result of this, a hybrid ML model including ANN and RF in the form of a tree structure to classify the level of pistachio adulterated images was built in this study.