Estimation of maturity-related components in edamame pods using a dual-input CNN with color and fluorescence images

Abstract

To enable rapid evaluation of maturity-related quality attributes of edamame, including moisture content and fructose content, a dual-input Convolutional Neural Network (CNN) model was developed using color images and 365 nm excitation fluorescence images as inputs. Color and fluorescence images were fed into independent CNN branches, respectively, and used to regress moisture content or fructose content. Hyperparameters were optimized using Bayesian optimization. Five-fold cross-validation results showed that EfficientNet-B0 achieved a Root Mean Square Error (RMSE) of 1.81 % for moisture content and an RMSE of 0.20 g/100 g DW for fructose content. Ablation analysis revealed that color images contributed predominantly to moisture content estimation, whereas fluorescence images contributed predominantly to fructose content estimation.