GreenFruitDetector: Lightweight green fruit detector in orchard environment

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Green fruit detection is challenging due to camouflage with foliage.
• Existing models often struggle with feature extraction and multi-scale object detection in orchards.

Purpose of the Study:

• To develop a lightweight and efficient model for accurate green fruit detection.
• To improve feature extraction, especially under occlusion, and enhance detection of small/distant fruits.

Main Methods:

• An improved YOLO v8 architecture incorporating Deformable Convolution and MCAG-DC (Multi-path Coordinate Attention Guided Deformer Convolution).
• A Fusion-neck structure for integrating multi-scale spatial information.
• A novel detection head designed for multi-scale object detection.
• Channel pruning techniques to reduce model size and computational cost.

Main Results:

• Achieved accuracies of 94.5% (Korla Pear), 84.4% (Guava), and 85.9% (Green Apple).
• Demonstrated accuracy improvements of 1.17%, 1.1%, and 1.77% over the baseline model.
• Increased mAP@0.5 by up to 6.5% and recall by up to 1.97%.
• Reduced model size, parameters, and FLOPs by 50%, 55%, and 44%, respectively.

Conclusions:

• The GreenFruitDetector model significantly enhances green fruit detection accuracy and efficiency.
• The architectural improvements effectively address challenges like color similarity and occlusion.
• The optimized model offers a practical solution for precision agriculture applications.