Distilled vision transformers with CNN fusion for robust cashew apple maturity prediction

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Cashew apples are highly nutritious but have a short shelf life due to their delicate nature.
• Accurate maturity grading is crucial for optimizing cashew apple post-harvest handling, storage, and market value.
• Current grading methods may not be efficient or scalable for the agricultural industry.

Purpose of the Study:

• To develop an efficient and accurate system for cashew apple maturity grading using artificial intelligence.
• To leverage knowledge distillation and lightweight transformer models for improved performance with limited data.
• To create a computationally efficient model suitable for real-time agricultural applications.

Main Methods:

• A lightweight vision transformer (ViT) student model was trained using multi-granular knowledge distillation (KD) from a stronger teacher model.
• The distillation process incorporated soft-label supervision, attention transfer, and token-level feature regression.
• Auxiliary lightweight models (MobileNet, ConvNeXt, EdgeNeXt) were used in an ensemble approach for enhanced predictions.

Main Results:

• The proposed ensemble model (ViT-KD with EdgeNeXt) achieved 90% accuracy on the test split.
• Stratified fivefold cross-validation demonstrated a mean accuracy of 86.89% ± 2.89%, indicating stable generalization.
• The distilled ViT model achieved real-time inference at 8.79 ms per image, outperforming conventional CNN baselines.

Conclusions:

• Knowledge-distilled lightweight transformers are effective for data-efficient maturity grading of cashew apples.
• The developed system offers a computationally efficient and accurate solution for agricultural applications.
• This approach has the potential to significantly improve post-harvest management and reduce food waste.