Vision-based Postharvest Analysis of Musa Acuminata Using Feature-based Machine Learning and Deep Transfer Networks

Abstract

Traditional practice of classifying postharvest crops creates variability in quality assessment due to human-related limitations including individual disparities in visual recognition. As a solution, computer vision approach was adapted. This study aims to classify the native banana fruit to South and Southern Asia (Musa acuminata) using the image-based deep transfer networks of ResNet101, MobileNetV2, and InceptionV3, and machine learning algorithms, including classification tree (CTree), Naïve Bayes algorithm (NB), k-nearest neighbors (KNN) and support vector machine (SVM). A total of 1,164 images, derived from 194 banana tier subjects with different orientations were utilized. Color channel thresholding in CIELab(L*a*b*) color space was applied to extract the spectral (RGB, HSV, YCbCr, L*a*b*), textural (correlation, contrast, entropy, homogeneity, energy) and morphological (total area) features. These 18-feature vectors were further simplified into two most significant features (S and V) using combined neighborhood component analysis and principal component analysis (hybrid NCA-PCA). The length of the top middle finger of the banana tier was added to the features. The classification tree (CTree), regardless of feature set, was validated to have the best performance, with accuracy of 91.28% and inference time of 19.34seconds. In addition, NB, KNN and SVM models provided acceptable performance with 89.72%, 89.30%, and 89.36% accuracies, respectively. However, the deep transfer networks did not provide acceptable classification results (with ResNet101 having the highest accuracy of 50.01% among the networks used). Lastly, the proposed machine learning models served as a feasible approach in the postharvest classification of Musa acuminata.