A Novel Method for Measuring the Color of Edible Oil on the Lovibond Scale Based on Spectral Detection and Convolutional Neural Network

Abstract

Abstract. A fast, high-precision method for measuring the color of edible oil on the Lovibond scale was developed and validated. It involves two components: spectral detection and data processing. The former was used for measuring the edible oil’s transmission spectrum in the visible range. The latter included a logarithmic method and a convolutional neural network (CNN) for calculating the color value from the transmission spectrum. The logarithmic method converted the multiplicative combination of spectra to an additive combination, which greatly improved the performance. A CNN with three convolutional layers was developed using Tensorflow. Validation tests of the method using ten different oil types showed a maximum error of 0.5, average error of 0.32, and error variance of 0.04. Thus, the proposed method can meet the needs of the edible oil industry, and it has great application potential. Keywords: CNN, Color of edible oil, Convolutional neural network, Lovibond scale, Spectral detection.