Machine learning-based pH quantification from test strip images using multiple color spaces

Abstract

Recent advances in colorimetric analysis have focused on improving measurement accuracy through the integration of smartphone cameras and machine learning. Particular focus has been placed on enhancing predictive accuracy by combining variables from multiple color spaces. In this study, we developed a model incorporating multiple color spaces using images of universal pH test strips. For model selection, we focused on localized errors and used the mean absolute error for each pH value as an evaluation metric. As a result, by emphasizing localized errors, we were able to build a model with better generalization performance than the model selected on the basis of overall error. We implemented this model in a real-time pH measurement demo web application.