Analysis of Convolutional Neural Network Models for Classifying the Quality of Dried Chili Peppers (Capsicum Annuum L)

In this paper, an analysis of convolutional neural network (CNN) models to classify the quality of dried chili pepper is described. The classifier models can determine the categories of a set of images that could be encountered in a sorting machine, such as “Extra”, “First class”, and “Second class” which correspond to different qualities of dried chili peppers. Additionally, two more classes were added as “Trash” and “Empty” which corresponds to cases that could occur in a sorting machine. To determine the best model for image classification, a set of state-of-the-art architectures were compared from the Torchvision library, including ResNet, ResNeXt, Wide ResNet, EfficientNet, and RegNet. The models were trained using feature extraction on the transfer learning approach, and were evaluated using cross-validation method and various advanced metrics such as Precision, Recall, Specificity, F1-score, Geometric mean, and Index of Balanced Accuracy. The results of the cross-validation process indicate that ResNet-152 is the best CNN model for implementation in a sorting machine, with a mean validation accuracy of 95.04%. By using this model, agricultural producers can ensure that their products are sorted according to international standards.