A Comparative Analysis of YOLOv5 and YOLOv8 for Watermelon Sweetness Prediction Based on Their Field-Spot

Abstract

Despite its popularity and high demand in the market, determining the sweetness and ripeness of watermelon remains difficult. Traditionally, ripe and sweet watermelon is identified by visual inspection of its weight, tapping sound, and visual characteristics such as color and field spot, which indicate its maturity time on the vine. However, this inspection method is frequently viewed as subjective, time-consuming, and prone to error. The objective of this research is to use an object identification algorithm to classify the sweetness level of watermelon based on its field spot, leading to a more precise and consistent evaluation. The architectures used in this study were YOLOv5 and YOLOv8. Both models were developed using 333 watermelon image data, which were categorized into two categories: sweet and non-sweet. The evaluation is carried out by contrasting the performance of both models concerning the values of precision, recall, and mean absolute performance (mAP 50 and mAP 50-90). The findings indicated that the YOLOv5 outperformed the YOLOv8 in determining the sweetness level of the watermelon, with mAP 50 values of 85.3\% and 81.9\%, respectively. Moreover, in contrast to YOLOv5, YOLOv8 necessitated an extensive training period.