Durum Wheat Classification Using Feature Selection, Bayesian Optimization and Support Vector

Wheat is the primary component of the majority of everyday food items, and acquiring high-quality wheat grains is a crucial concern for the production of food products. Recognizing the types of durum wheat is vital during processing in food-processing facilities. A dataset that included two varieties of durum wheat and extraneous substances was gathered. The objective of this study was to identify the minimum number of features from a pool of 236 morphological, color, wavelet, and gaborlet features, which can yield the highest accuracy with minimal difference between validation and test accuracy for two kinds of durum wheat: starchy durum wheat and vitreous durum wheat and foreign elements. This study proposes a machine learning approach to identify the optimal set of features for distinguishing between starchy, vitreous durum wheat, and foreign elements. This approach comprises feature selection, optimization, and classification. First, five feature selection techniques, MRMR, ChiSquare, Relief, ANOVA, and Kruskal-Wallis with SVM, were evaluated for identification of durum wheat. After conducting the analysis, it was found that out of the 236 features, a set of 50 features yielded significant performance. However, it also suffers of decreasing 2-3% decrease in accuracy. To compensate for this, a Bayesian optimization technique was introduced with SVM, which achieved a validation accuracy of 99.8% and test accuracy of 99.6%. This methodology helps to identify vitreous durum wheat in the food-processing chain.