Automated Bell Pepper Quality Assessment: Robotic Gripper Sorting System with Transfer Learning

Abstract

Sorting is an activity during post-harvest that separates fresh produce depending on certain parameters. If this activity is manually done, it is time-consuming and sometimes inconsistent. The marketability of fruits and vegetables often relies on customers’ standards and satisfaction. When these standards are not met, this will result in food wastage in the long run. In this study, the researchers aim to develop a sorting system using three (3) transfer learning algorithms with a robotic gripper application – which has not been majorly explored in previous studies. Moreover, this study also intends to aid bell pepper retailers in preventing food loss due to unsatisfied customer preferences. The process starts with image acquisition for data gathering. The collected data is subjected to data splitting for training and testing. Three pre-trained algorithms were used namely; VGG-16, Resnet50, and GoogleNet. Each of which undergone three train-test splits of; 70-30%, 75-25%, and 80-20% to see their accuracy. VGG-16 obtains an accuracy of 98.38% for both 70-30% and 75-25% train-test split. GoogleNet on the other hand, has the highest accuracy on 80-20% split with 97.84%. ResNet50 has the lowest accuracy having 90.23% for train-test split of 75-25%.