Related Experiment Video
Updated: Aug 5, 2026
11:49
Cereal Crop Ear Counting in Field Conditions Using Zenithal RGB Images
Published on: February 2, 2019
A Comparative Benchmark of Real-Time Detectors for Canopy Image-Based Blueberry Detection Toward Precision Orchard
Xinyang Mu1, Yuzhen Lu1, Boyang Deng1
1Department of Biosystems and Agricultural Engineering, Michigan State University, East Lansing, MI 48824, USA.
Sensors (Basel, Switzerland)
|July 28, 2026
Summary
This study benchmarks AI object detectors for blueberry farming, finding RT-DETR models offer top accuracy. Semi-supervised learning further boosts performance, enhancing computer vision applications in agriculture.
Area of Science:
- Agricultural technology
- Computer vision
- Artificial intelligence
Background:
- Computer vision using artificial intelligence (AI) presents a valuable tool for blueberry cultivation, aiding tasks like maturity assessment and yield estimation.
- Challenges in blueberry detection include variable lighting, occlusions, and motion blur, necessitating robust AI models.
- Deep learning object detectors require large, diverse datasets for effective real-world deployment, balancing accuracy, speed, and memory constraints.
Purpose of the Study:
- To conduct a comparative benchmark analysis of advanced real-time object detectors for blueberry detection.
- To evaluate the performance of YOLO (You Only Look Once) and RT-DETR (Real-Time Detection Transformers) model families on a new, large blueberry dataset.
- To assess the impact of semi-supervised learning (SSL) on improving fruit detection accuracy.
Main Methods:
- Developed a large dataset of 661 smartphone-captured canopy images with 85,879 manually annotated blueberry instances.
- Benchmarked 36 variants of YOLO (v8-v12) and RT-DETR (v1-v2) object detectors on the curated dataset.
- Fine-tuned selected models using Unbiased Mean Teacher-based semi-supervised learning (SSL) with 1644 unlabeled cross-source images.
Main Results:
- RT-DETRv2-X achieved the highest accuracy (mAP@50 of 93.6%) among evaluated RT-DETR models, while YOLOv12m led the YOLO family (mAP@50 of 93.3%).
- Mid-sized models demonstrated a favorable balance between accuracy and inference speed.
- SSL fine-tuning improved accuracy by up to 2.0%, with RT-DETR-v2-X reaching a top mAP@50 of 95.5%.
Conclusions:
- Advanced real-time object detectors, particularly RT-DETR, show high potential for blueberry detection tasks.
- Semi-supervised learning effectively leverages cross-domain unlabeled data to enhance fruit detection performance.
- The publicly available dataset and software facilitate future research and practical implementation of AI in blueberry farming.
