A task-specific architecture with multi-scale attention and shape-aware loss for strawberry phenophase recognition in...

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Accurate strawberry phenological stage detection is crucial for intelligent agriculture but challenging in complex field environments.
• Existing methods struggle with recognizing small strawberry targets and subtle visual cues across growth stages.

Purpose of the Study:

• To develop a novel, end-to-end lightweight detection architecture (HCMS-Net) for precise strawberry phenological perception.
• To enhance the model's ability to identify small targets and capture long-range dependencies for continuous phenological changes.

Main Methods:

• Utilized a Residual Efficient Layer Aggregation Network (R-ELAN) backbone with Multi-Scale Convolutional Attention (MSCA).
• Incorporated hypergraph convolution and a Mixed Aggregation Network (MANet) for feature fusion.
• Employed a Conv2Former module in the detection head and a Shape-Normalized Wasserstein Distance (Shape-NWD) loss function.

Main Results:

• HCMS-Net achieved a mean average precision (mAP) of 94.9% and an F1-score of 90.0%.
• Demonstrated superior performance over ten mainstream detectors, including RT-DETR and YOLO variants, with significantly fewer parameters.
• Heatmaps confirmed precise attention focus across all five phenological stages, effectively ignoring background noise.

Conclusions:

• HCMS-Net offers high accuracy and efficiency for strawberry phenological period detection.
• The model supports advancements in selective harvesting and intelligent agricultural management systems.
• The architecture effectively addresses challenges in small target recognition and complex environmental perception.