Detection and Maturity Classification of Dense Small Lychees Using an Improved Kolmogorov-Arnold Network-Transformer.

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Accurate lychee detection and maturity classification are vital for optimizing harvest yields.
• Dense fruit clusters present significant challenges for ripeness assessment due to limited visibility and training data.

Purpose of the Study:

• To develop an efficient and accurate deep learning model for lychee detection and ripeness classification in complex on-tree cluster scenarios.
• To address the limitations of existing models in handling dense fruit arrangements and small object detection.

Main Methods:

• Proposed a novel GhostResNet (GRN)-KAN-Transformer model integrating GhostResNet modules for efficient feature extraction and Kolmogorov-Arnold Networks (KAN) for enhanced non-linear mapping.
• Introduced a large-scale layer for improved small object sensitivity and a multi-layer cross-fusion attention (MCFA) module for deeper hierarchical feature integration.

Main Results:

• The GRN-KAN-Transformer model achieved significant reductions in GFLOPs (8.84%) and parameters (11.24%) compared to the baseline.
• Achieved high mean Average Precision (mAP) scores of 94.7% (mAP50) and 58.4% (mAP50-95).
• Demonstrated superior performance against established models like YOLOv8n, YOLOv12n, CenterNet, and EfficientNet.

Conclusions:

• The GRN-KAN-Transformer model offers a computationally efficient yet highly accurate solution for lychee detection and ripeness classification.
• The model's effectiveness in dense clusters suggests potential for broader application in precision agriculture for fruit monitoring and management.