A comparative study of convolutional neural networks and traditional feature extraction techniques for adulteration detection in ground beef

The presence of adulteration in meat products, including minced meat, is a serious concern in many parts of the world. Therefore, notable efforts have been made to find fast, non-destructive and efficient methods to detect adulteration in minced meat. In this research, thermal imaging was investigated to detect adulteration of ground beef in two data sets that included sheep lung and chicken gizzard as impurities. In order to identify the most appropriate feature extraction algorithm and classify samples having various levels of adulteration, Local Binary Pattern (LBP), Gray Level Co-occurrence Matrixes (GLCM) and Gabor filter were compared. Convolutional Neural Network (CNN) was also used to extract features and classify images. In order to evaluate these algorithms, the following criteria were utilized: accuracy, precision, recall, specificity and F-score. Results showed that for both datasets, the best performance was obtained using the Gabor filter while the weakest performance was related to the LBP algorithm. However, CNN, with a total accuracy of over 99 % in both data sets, was found to surpass the other methods and is recommended as the best approach for analyzing thermal images of ground beef adulterated with avian and ovine offal. This shows the superiority of CNN algorithm over machine learning algorithms in identifying adulteration in minced meat. The experimental results and the associated data analysis presented here show the appropriate use of thermography in identifying meat fraud, which can be suitable for online applications.