Image based beef and lamb slice authentication using convolutional neural networks

Abstract

Meat adulteration affects customers and the market. Existing meat authentication methods usually rely on special devices, and thus are limited to professional use only. Fake lamb or beef slices made from duck and fat appear in some Chinese hotpot restaurants. This study present a customer–conducted method for detecting such adulteration. The method takes a single image as input, and authenticates it by using the textural features of meat slices. The authentication is rapid and non–destructive. The only device required is a mobile phone with a camera. A lightweight (thus high–efficient) convolutional neural network architecture called MTx–Net was built for this task. Fourteen convolutional layers in four blocks were used for extracting valid visual features. Techniques like residuals, depth–wise convolution, dropout and batch normalization were employed in the neural network. For training and testing the neural network, 77956 meat images were collected using 225Kg of meats. This method achieves 99.38% and 98.20% accuracy on lamb and beef slice authentication, respectively.