KD-SSGD: knowledge distillation-enhanced semi-supervised germination detection.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Precision agriculture demands accurate seed germination detection for crop monitoring and variety selection.
• Fully supervised methods require extensive annotated datasets, which are costly and time-consuming in agricultural settings.

Purpose of the Study:

• To develop an efficient semi-supervised learning framework for seed germination detection that minimizes reliance on labeled data.
• To introduce a novel knowledge distillation approach that enables end-to-end training without a pre-trained teacher model.

Main Methods:

• A teacher-student architecture incorporating a lightweight distilled student branch.
• Key modules include Weighted Boxes Fusion (WBF) for pseudo-label optimization, Feature Distillation Loss (FDL) for knowledge transfer, and Branch-Adaptive Weighting (BAW) for training stability.

Main Results:

• Achieved 47.0% mAP on the Maize-Germ dataset using only 1% labeled data, outperforming existing semi-supervised methods.
• Demonstrated strong performance on the Three Grain Crop dataset, with mAP reaching up to 76.1% at 10% labeled data.
• Showcased robust cross-crop generalization capabilities and effective knowledge transfer under limited supervision.

Conclusions:

• The KD-SSGD framework provides high-quality pseudo-labels and stable, high-precision detection with minimal labeled data.
• This approach offers an efficient and scalable solution for intelligent agricultural perception and automated crop monitoring.
• The method significantly reduces the annotation burden, making advanced computer vision techniques more accessible for agricultural applications.