Application of Deep Learning for Automated Peach Classification: A Study Based on ResNet Architectures

Abstract

This study evaluates the performance of various ResNet architectures for classifying peaches as “healthy” or “damaged”. A dataset of 3 370 images was used, with data-augmentation techniques applied to enrich the training set. Transfer learning was performed using pre-trained ResNet models, with stochastic gradient descent (SGD) adopted as the optimisation algorithm. Performance was assessed using accuracy, precision, recall and F1 score. ResNet-50 emerged as the most effective architecture, achieving a mean accuracy of 95.96 % and outperforming other models, including ResNet-18, ResNet-34, ResNet-101 and ResNet-152. The results demonstrate the potential of deep-learning techniques to improve peach-sorting processes, thereby reducing post-harvest losses and enhancing quality control in the agricultural sector.