Evaluating CNN Models and Optimization Techniques for Quality Classification of Dried Chili Peppers (Capsicum annuum L.)

Abstract

This paper analyzes Convolutional Neural Network (CNN) models for classifying dried chili pepper quality. The models categorize images into five categories: “Extra”, “First Class”, “Second Class”, “Trash”, and “Empty”, each representing different qualities and scenarios in a sorting machine. We compared architectures from the Torchvision library, including ResNet, ResNeXt, Wide_ResNet, and RegNet using Transfer Learning (TL) in a feature extraction approach. All models employ residual blocks, an innovative technique enhancing deep learning performance. The models were evaluated using crossvalidation and metrics such as Precision, Recall, Specificity, F1-score, Geometric_mean, Index of Balanced Accuracy, and the Matthews Correlation Coefficient. They were trained using SGD, Adagrad, and Adam optimizers. Our findings suggest that ResNet-152, trained with the Adagrad optimizer, achieved the highest mean validation accuracy of 96.62%. The selected model can assist agricultural producers in classifying their products according to international standards.