Zhang Chunjuan, Zhang Dequan, Su Yuanyuan, Zheng Xiaochun, Li Shaobo, Chen Li
Institute of Food Science and Technology, Chinese Academy of Agricultural Sciences, Key Laboratory of Agro-Products Quality and Safety Control in Storage and Transport Process, Ministry of Agriculture and Rural Affairs, Beijing 100193, China.
School of Food and Wine, Ningxia University, Yinchuan 750021, China.
Foods. 2022 Nov 21;11(22):3732. doi: 10.3390/foods11223732.
To realize the real-time automatic identification of adulterated minced mutton, a convolutional neural network (CNN) image recognition model of adulterated minced mutton was constructed. Images of mutton, duck, pork and chicken meat pieces, as well as prepared mutton adulterated with different proportions of duck, pork and chicken meat samples, were acquired by the laboratory's self-built image acquisition system. Among all images were 960 images of different animal species and 1200 images of minced mutton adulterated with duck, pork and chicken. Additionally, 300 images of pure mutton and mutton adulterated with duck, pork and chicken were reacquired again for external validation. This study compared and analyzed the modeling effectiveness of six CNN models, AlexNet, GoogLeNet, ResNet-18, DarkNet-19, SqueezeNet and VGG-16, for different livestock and poultry meat pieces and adulterated mutton shape feature recognition. The results show that ResNet-18, GoogLeNet and DarkNet-19 models have the best learning effect and can identify different livestock and poultry meat pieces and adulterated minced mutton images more accurately, and the training accuracy of all three models reached more than 94%, among which the external validation accuracy of the optimal three models for adulterated minced mutton images reached more than 70%. Image learning based on a deep convolutional neural network (DCNN) model can identify different livestock meat pieces and adulterated mutton, providing technical support for the rapid and nondestructive identification of mutton authenticity.
为实现掺假羊肉末的实时自动识别，构建了掺假羊肉末的卷积神经网络（CNN）图像识别模型。通过实验室自建的图像采集系统获取羊肉、鸭肉、猪肉和鸡肉块的图像，以及用不同比例的鸭肉、猪肉和鸡肉掺假的羊肉样本图像。所有图像中，不同动物种类的图像有960张，鸭肉、猪肉和鸡肉掺假的羊肉末图像有1200张。另外，重新采集了300张纯羊肉以及鸭肉、猪肉和鸡肉掺假羊肉的图像用于外部验证。本研究比较分析了AlexNet、GoogLeNet、ResNet-18、DarkNet-19、SqueezeNet和VGG-16这六种CNN模型对不同畜禽肉块及掺假羊肉形状特征识别的建模效果。结果表明，ResNet-18、GoogLeNet和DarkNet-19模型学习效果最佳，能更准确地识别不同畜禽肉块及掺假羊肉末图像，三种模型的训练准确率均达到94%以上，其中最优的三种模型对掺假羊肉末图像的外部验证准确率达到70%以上。基于深度卷积神经网络（DCNN）模型的图像学习能够识别不同畜禽肉块及掺假羊肉，为羊肉真伪的快速无损识别提供技术支持。
