Design and Implementation of a Computer Vision System Using Convolutional Neural Networks for Sustainable Quality and Maturity Assessment of Bell Peppers

Destructive testing is usually used for sorting and managing produce in a retail environment, which results in further damage and the misclassification of sensitive products. The quality of vegetables and fruits like bell peppers is reflected in their external features. Therefore, this research aims to develop a non-destructive method that uses a computer vision system and machine learning algorithms that can automatically determine the quality and maturity of bell peppers in real-time applications. This study implements two Convolutional Neural Network (CNN) machine learning models: YOLOv8 and EfficientNetV2 B0. Additionally, automated rollers are added to rotate multiple bell peppers, allowing the camera to capture different perspectives for more detailed classification. The YOLOv8 was used to obtain multiple images of bell peppers, which EfficientNetV2 B0 then classified. The bell peppers were categorized into four categories: (1) Ripe with good condition, (2) Unripe with good condition, (3) Ripe with bad condition, and (4) Unripe with bad condition. The accuracy of the device was tested in two supermarkets. In the first supermarket, the automated classification obtained an accuracy of 93.2 %, while manual classification obtained an accuracy of 90.6 %. For the second supermarket, the automated classification obtained an accuracy of 91.4 %, and the manual classification accuracy was 89.7 %. Overall, the automated classification achieves an accuracy of 92.3 %, surpassing the overall manual classification accuracy of 90.1 %.