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

Abstract

Conventional methods for determining protein and oil content in rapeseed are often time-consuming, labor-intensive, and costly. In this study, a mobile application was developed using an optimized deep learning method for low-cost, non-destructive and real-time prediction of protein and oil content in rapeseed by inputting rapeseed images. Among the tested models, FasterNet-L showed the optimal performance, with predicted coefficients of determination (R_p²) of 0.9366 for oil content and 0.8828 for protein content. The mean square error of prediction (RMSEP) was 0.6982 and 0.6498, and the residual predictive deviation (RPD) was 3.88 and 2.92 for oil and protein content, respectively. Furthermore, three pruning methods were employed, and neural pruning via growth regularization proved to be the most effective, with a 13.18 % improvement in prediction speed and a 15.79 % reduction in model size. Finally, this method can be expanded and applied to other oilseed crops for rapid quality identification and detection.