Segmentation of Mandarin Oranges With Deep Learning for Postharvest Quality Analysis

ABSTRACT This study contributes to advancements in agricultural technology and food security by employing YOLO‐based deep learning techniques on Raspberry Pi for fruit segmentation and quality assessment. The research utilizes two object detection architectures, namely YOLOv5 and YOLOv7, to segment mandarin oranges using state‐of‐the‐art deep learning architectures. The dataset encompasses three classes: over‐ripe, ripe and unripe mandarin oranges. Polygon techniques are applied during preprocessing to ensure precise segmentation. YOLOv5 demonstrates impressive performance, achieving a recall score of 0.984 and a precision score of 0.977 for both box detection and mask segmentation, with mean average precisions (mAPs) of 0.994 at an IoU threshold of 0.5. Conversely, YOLOv7 exhibits relatively lower performance, with box detection and mask segmentation accuracies of 0.925 and recall scores of 0.972, yielding mAPs results of 0.992 for box detection and mask segmentation at an IoU threshold of 0.5. Across a broader IoU spectrum from 0.5 to 0.95, YOLOv7 achieves mAPs of 0.882 for box detection and 0.87 for mask segmentation. Utilizing a YOLO‐based deep learning model has led to the development of an accurate and economical system for categorizing oranges based on their maturity levels.