Lightweight pear detection in unstructured orchards via selective information propagation

Area of Science:

• Computer Vision
• Agricultural Robotics
• Machine Learning

Background:

• Accurate pear detection is crucial for automated harvesting and orchard management.
• Challenges include similar coloration to foliage, spherical shape, and fruit occlusion in clusters.
• Existing lightweight detectors struggle with accuracy in complex orchard environments.

Purpose of the Study:

• To develop a lightweight and accurate pear detection framework for unstructured orchards.
• To address challenges like color similarity, shape, and occlusion using selective information propagation.
• To enable efficient deployment on embedded agricultural platforms for real-time applications.

Main Methods:

• Proposed a lightweight detection framework based on selective information propagation.
• Implemented modules for global-local context modeling, adaptive feature transformation, and detail-preserving fusion.
• Utilized an adaptive IoU loss function for small and occluded fruits.
• Developed a self-built Orchard Pear dataset for training and evaluation.

Main Results:

• Achieved 95.2% mAP@50 and 54.6% mAP@50:95 on the Orchard Pear dataset.
• The model has only 2.56 million parameters and 5.60 GFLOPs.
• Demonstrated consistent improvements on public Minne Apple and Mango datasets.
• Validated real-time inference capabilities on embedded platforms.

Conclusions:

• Selective feature representation, fusion, and optimization are key for lightweight fruit detection in complex scenes.
• The proposed framework offers a viable solution for real-time robotic harvesting and orchard perception.
• The method effectively handles challenges posed by natural orchard environments.