Classification of Meat Freshness Based on Deep Learning Using Data from Diffuse Reflectance Spectroscopy

Abstract

Met-myoglobin is a major component related to meat discoloration, and it gradually accumulates over time after the meat is slaughtered. Recently, studies have been conducted to observe the changes in the composition of met-myoglobin in the meat along with its storage time using Diffuse Reflectance Spectroscopy(DRS). DRS is an optical technique that is simple and can estimate the composition of chromophores without damaging the sample. DRS illuminates the light to the sample and obtains a diffuse reflectance spectrum, and the composition of the chromophore is calculated through an inverse modeling process. However, since DRS requires high resolution and complicated fitting process, it is difficult to apply DRS to the mobile environment. Therefore, the purpose of our study is to classify the freshness of meat by extracting features from low spectral resolution diffuse reflectance spectrum by using the deep learning model. Converting the diffuse reflectance spectrum into 1D-vector form without any preprocessing and putting it into 1D convolutional neural network(CNN) showed 96.65(±1.17)% accuracy. To improve the generality of the model, a data augmentation was used. As a result, the accuracy of the using diffuse reflectance spectrum was 88.18(±3.48)%. To consider the applicability at low-resolution spectrometer, the diffuse reflectance spectrum was down-sampled 5, 10, 30 and 50 times. Even though down-sampled to 50 times, accuracy was higher than 88%, and it demonstrated its applicability in mobile environment.