This study was aimed at evaluating the effect of spontaneous lacto-fermentation of carrot slices on flesh structure using different machine learning approaches. The textures computed from digital images of lacto-fermented and fresh carrot slices were compared using neural networks and other algorithms from different groups. In the case of Multilayer Perceptron, accuracies for training, testing and validation were considered. For some of the networks, lacto-fermented and fresh samples were completely distinguished. The accuracies for training, testing and validation were equal to 100%. For models built using other algorithms (LDA (Linear Discriminant Analysis), Multi Class Classifier, LMT (Logistic Model Tree), KStar, Naive Bayes, PART), the following metrics were used for the evaluation of model effectiveness: accuracies, time taken to build model, Kappa statistic, mean absolute error, root mean squared error, PRC (Precision-Recall) Area, ROC (Receiver Operating Characteristic) Area, MCC (Matthews Correlation Coefficient), F-Measure, Recall, Precision, FP (False Positive) Rate and TP (True Positive) Rate. The most satisfactory results were obtained for the LDA. The lacto-fermented and fresh carrot slices were distinguished with an average accuracy of 99%, low values of errors (mean absolute error: 0.0117, root mean squared error: 0.1014) and FP Rate (0.010). The weighted averages of other metrics were greater than or equal to 0.98 (Kappa statistic: 0.98, PRC Area: 0.987, ROC Area: 0.991, MCC: 0.980, F-Measure: 0.990, Recall: 0.990, Precision: 0.990, TP Rate: 0.990). The obtained results demonstrated the usefulness of different machine learning approaches to the evaluation of the effect of fermentation on changes in the carrot flesh structure.
