Blueberry Segmentation and Ripeness Assessment
Deep learning-based detection and classification of blueberries for pre-harvest ripeness assessment (FONDECYT 171-2020)
Project Overview
This government-funded research project (FONDECYT 171-2020) developed an artificial vision system for automated blueberry detection and ripeness classification in agro-industrial environments. This system enables automated fruit sampling during pre-harvest operations, addressing the labor-intensive nature of manual ripeness assessment in large blueberry fields.
Peru has become a major contributor to the international blueberry market, exporting to the UK, US, and China. However, traditional ripeness assessment methods remain time-consuming, labor-intensive, and prone to human error from worker fatigue. Our system automates this process using deep learning techniques.
System Architecture
The system consists of two main modules:
Object Detection Module: YOLOv8 detects and segments blueberries in images, outputting bounding boxes and instance masks as Regions of Interest (ROIs).
Classification Module: DINOv2 extracts embeddings from each ROI, which are then classified by an MLP head into one of five ripeness stages.
Ripeness Stages
| Stage | Name | Description |
|---|---|---|
| 1 | Green | Immature, fully green |
| 2 | Cream | Early color change |
| 3 | Blush | Intermediate ripeness |
| 4 | Pink | Near-ripe |
| 5 | Blue | Fully mature |
Data Collection
Images were captured using a ZED 2i stereo camera at agro-industrial farms in Trujillo, Peru. The detection dataset contains 200 images with 1,708 labeled blueberry instances, while the classification dataset comprises 900 ROIs balanced across the five ripeness classes.
Results
The system achieved 93.70% classification accuracy across five ripeness stages, with the YOLOv8 Medium backbone achieving 91.9% mAP@50 for instance segmentation and 92.1% mAP@50 for object detection. The confusion matrix shows the classification performance across all five ripeness stages.
Publications
- Artificial Vision Strategy for Ripeness Assessment of Blueberries on Images Taken During Pre-harvest Stage in Agroindustrial Environments Using Deep Learning Techniques (2023)
- P. B. Cubas Muñoz, R. J. Huaman Kemper, and S. R. Prado Gardini
- IEEE XXX International Conference on Electronics, Electrical Engineering and Computing (INTERCON)
- DOI: 10.1109/INTERCON59652.2023.10326058
Workshop Presentations
- Deep Learning Approach for Accurate Pre-Harvest Blueberry Ripeness Classification (2023)
- P. Cubas, R. J. Huaman Kemper, and S. R. Prado Gardini
- Workshop on Robotics in Agriculture: Present and Future of Agricultural Robotics and Technologies
- IEEE/RSJ IROS 2023, Detroit, USA
- Workshop Website
Acknowledgments
This research was funded by FONDECYT (Fondo Nacional de Desarrollo Científico, Tecnológico y de Innovación Tecnológica) under project 171-2020-FONDECYT and conducted at the Laboratorio de Investigación Multidisciplinaria (LABINM) at Universidad Privada Antenor Orrego (UPAO), Trujillo, Peru.