Classification and quantification of minced mutton adulteration with pork using thermal imaging and convolutional neural network

Abstract

A novel and reliable method to classify and quantify the adulterated minced mutton is proposed in this study based on thermal imaging combined with convolutional neural network (CNN). Firstly, thermal videos of 35 pure mutton samples, 35 pure pork samples, and 175 adulterated mutton samples (minced mutton mixed with pork at different levels: 10-20-30-40-50%) during continuous heating were collected. Secondly, thermal images at the faster heating rate stage were extracted, and their information at regions of interest was obtained for building the qualitative and quantitative CNN models. When the models were developed, learning rate and mini batch of models were selected by parameter comparison. Finally, the optimal qualitative classification model using Softmax classifier was determined to classify pure mutton samples, adulterated mutton samples and pure pig samples, and the optimal quantitative prediction model using regression function was determined to predict the pork proportion of adulterated mutton mixed with minced pork. The results showed that, the accuracy of validation and test sets of the qualitative CNN model was 99.97% and 99.99%, respectively; the determination coefficient (R2), root mean square error (RMSE) and relative prediction deviation (RPD) of validation and test sets of the quantitative CNN model were 0.9933, 0.0251, 12.2487 and 0.9933, 0.0252, 12.2387, respectively. Therefore, thermal imaging combined with CNN has achieved good results in qualitative classification of different samples and quantitative prediction of adulterated proportion. Due to its economy and convenience, this method has great application potential in the detection and supervision of adulterated food.