Soft-labeling approach along with an ensemble of models for predicting subjective freshness of spinach leaves

In the previous studies, machine learning models generally used the majority vote or average value to predict agricultural product freshness. However, freshness evaluation is subjective owing to inherent differences in individuals' perceptions, and it is difficult to reach a consensus. In this study, we predicted distributions of spinach leaf freshness considering the uncertainty of human subjective. We used a dataset consisting of four classes for 1,045 images with 12 annotations from 12 panels. Hard-labeling approaches with probabilistic output and a softlabeling approach along with an ensemble of models were used to predict freshness distributions. The similarity between human freshness evaluation and the output distribution from the models were compared. Using ResNet152 (V1) with multi-output multi-class (MOMC) obtained the best result. Two metrics, histogram intensity of 0.76 and Kolmogorov-Smirnov value of 0.23, indicate high performance. Additionally, the ensemble methods, MOMC, predicted the mean of freshness values with the coefficient of determination of 0.74 and root mean square error of 0.34. The models incorporating human uncertainty provides realistic predictions, which are similar to the subjective levels of freshness evaluation.

Automated summary

In previous studies, machine learning models typically used majority vote or average value to predict agricultural product freshness. However, this study considers the subjective nature of freshness evaluation and predicts distributions of spinach leaf freshness by incorporating human uncertainty. The models achieved high performance, with similarity metrics between human freshness evaluation and output distribution indicating realistic predictions.