YOLOv4-Driven Appearance Grading Filing Mechanism: Toward a High-Accuracy Tomato Grading Model through a Deep-Learning Framework

Abstract

In traditional agricultural quality control, agricultural products are screened manually and then packaged and transported. However, long-term fruit storage is challenging in tropical climates, especially in the case of cherry tomatoes. Cherry tomatoes that appear rotten must be immediately discarded while grading; otherwise, other neighboring cherry tomatoes could rot. An insufficient agricultural workforce is one of the reasons for an increasing number of rotten tomatoes. The development of smart-technology agriculture has become a primary trend. This study proposed a You Only Look Once version 4 (YOLOv4)-driven appearance grading filing mechanism to grade cherry tomatoes. Images of different cherry-tomato appearance grades and different light sources were used as training sets, and the cherry tomatoes were divided into four categories according to appearance (perfect (pedicled head), good (not pedicled head), defective, and discardable). The AI server ran the YOLOv4 deep-learning framework for deep image learning training. Each dataset group was calculated by considering 100 of the four categories as the difference, and the total numbers of images were 400, 800, 1200, 1600, and 2000. Each dataset group was split into an 80% training set, 10% verification set, and 10% test set to overcome the identification complexity of different appearances and light source intensities. The experimental results revealed that models using 400–2000 images were approximately 99.9% accurate. Thus, we propose a new mechanism for rapidly grading agricultural products.