A Novel Curriculum Learning Training Strategy for Pomegranate Growth Stage Classification Using YOLO Models on Multi-Source Datasets for Precision Agriculture

Pomegranates are among the many vital crops generally believed to offer health quality and economically impact agriculture. Accurate detection and classification of the pomegranate growth stages enables fruit harvesting robots, resulting in yield optimization, supply chain, and market readiness. In this study, we propose that the YOLO object detection models adopt a novel curriculum learning approach to detect the growth stages of pomegranates using multi-source datasets. A combined dataset of 5700+ images is considered for categorizing pomegranates into bud, flowering, early-fruit, mid-growth, and maturity stages. In the present study, we have experimented with two deep learning-based object detection models, YOLOv5 and YOLOv7, to perform training using a multi-level curriculum learning approach. The proposed training strategy proves to show improvement over state-of-the-art work Zhao et al. (2024) in achieving higher detection accuracies towards all five classes. The efficiency of the curriculum learning strategy with YOLOv7 and YOLOv5 models achieved a mAP score of 92.2% with a precision of 90.1%, recall of 82.3%, and F1 score of 86.1%. By comparing the performance proposed curriculum learning-based YOLO object detection models with traditional learning-based YOLO models, it was revealed that the YOLOv5 model with curriculum learning shows a consistent improvement over another model. It is inferred that the proposed training strategy is valuable in increasing the efficacy of pomegranate fruit detection and enhancing precision agriculture activities.