Enhanced geometric properties prediction for carrots in motion using a Multi-task YOLO-based Linked Network (YL-Net)

With a growing shortage of manual labor in agriculture, there is an urgent need for efficient and automated solutions to address the complex challenges of inspecting and classifying agricultural products, such as carrots, which exhibit irregular shapes, occlusions, and high variability in geometric properties. This study introduces a novel Multi-task YOLO-based Linked Network (YL-Net) designed to estimate the geometric properties of carrots in motion, including width, length, volume, and mass. The proposed network integrates RGB-D input with decoupled multi-task learning to simultaneously perform instance segmentation and regression. Building upon our previous work, the enhanced framework presented herein achieves outstanding performance, with MAPE values below 2.5% for all estimated properties. When aggregating multi-view data from a rolling conveyor system, the accuracy further improves, yielding MAPE values below 2%. In terms of detection, the model demonstrates excellent performance, achieving a mean F1-score of 98.78% and an instance segmentation IoU of 89.25%. To evaluate its scalability, the system was deployed on an NVIDIA Jetson Orin Nano, where it achieved a real-time processing speed of 80 FPS. Beyond carrots, the proposed approach can be extended to inspect other agricultural products, such as potatoes and sweet potatoes, where geometric properties are essential for sorting and grading. This work provides a scalable and transferable solution for automated agricultural inspection, laying a robust foundation for broader applications in smart farming, industrial automation, and food quality control.