Developing Machine Vision in Tree-Fruit Applications-Fruit Count, Fruit Size and Branch Avoidance in Automated...

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Advancements in depth imaging hardware and 2D Convolutional Neural Networks (CNNs) are driving machine vision adoption.
• Mainstream models now often surpass application-specific architectures in performance.
• Publicly releasing training and test datasets is crucial for validating and re-evaluating research.

Purpose of the Study:

• To evaluate the effectiveness of mainstream CNN models for key mango orchard applications: fruit counting, fruit sizing, and branch avoidance.
• To quantify the impact of stochastic processes on model retraining.
• To assess the viability of edge computing devices for real-time agricultural tasks.

Main Methods:

• Utilized YOLOv8m and YOLOv9 models for object detection and segmentation tasks.
• Compared performance against benchmark models like MangoYOLO and Mask R-CNN.
• Developed and implemented a branch avoidance algorithm using YOLOv8m-seg.
• Ensured all training and test datasets are publicly available.

Main Results:

• YOLOv8m achieved 99.3% mAP50 for real-time mango fruit detection, outperforming MangoYOLO.
• YOLOv8 and v9 models exceeded Mask R-CNN in accuracy and inference speed for fruit and branch detection.
• YOLOv8m-seg demonstrated comparable accuracy to Mask R-CNN for fruit sizing with significantly reduced inference time.
• Low coefficient of variation (0.2%–2%) in mAP across repeated trainings indicates model stability.

Conclusions:

• Mainstream CNN models, particularly YOLOv8, offer superior performance and efficiency for automated mango harvesting tasks compared to previous benchmarks.
• The short inference times enable real-time implementation on edge computing devices, facilitating field adoption.
• Publicly available datasets and reproducible training practices enhance the reliability and re-evaluation of machine vision research in agriculture.