Lightweight deep learning model for embedded systems efficiently predicts oil and protein content in rapeseed

Area of Science:

• Agricultural Science
• Computer Science
• Biotechnology

Background:

• Traditional methods for analyzing rapeseed quality (protein and oil content) are inefficient, requiring significant time, labor, and cost.
• There is a need for rapid, non-destructive, and cost-effective techniques for quality assessment in the agricultural sector.

Purpose of the Study:

• To develop a mobile application utilizing an optimized deep learning model for real-time, non-destructive prediction of protein and oil content in rapeseed.
• To evaluate the performance of different deep learning models and pruning techniques for this application.

Main Methods:

• Image acquisition of rapeseed samples.
• Development and optimization of a deep learning model (FasterNet-L) for image-based quality prediction.
• Application of neural pruning techniques (neural pruning via growth regularization) to enhance model efficiency.

Main Results:

• The FasterNet-L model achieved high prediction accuracy, with R_p² values of 0.9366 for oil and 0.8828 for protein content.
• Neural pruning via growth regularization significantly improved prediction speed by 13.18% and reduced model size by 15.79%.
• The developed method demonstrated robust performance with low prediction errors (RMSEP) and good predictive capability (RPD).

Conclusions:

• A mobile application powered by deep learning offers a viable, low-cost, and efficient solution for real-time protein and oil content determination in rapeseed.
• The optimized deep learning approach, including model pruning, enhances prediction speed and reduces model size, making it practical for field applications.
• This technology has the potential for broad application in the rapid quality assessment of various oilseed crops.