Automated tea quality identification based on deep convolutional neural networks and transfer learning. (22nd February 2023)

• Record Type: Journal Article
• Authors: Zhang, Cheng; Wang, Jin; Lu, Guodong; Fei, Shaomei; Zheng, Tao; Huang, Bincheng
• Is Part Of: Journal of food process engineering. Volume 46:Number 4(2023)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2023-02-22
• Subjects: computer vision -- deep learning -- fine‐tune -- tea quality identification -- transfer learning; Food industry and trade -- Periodicals; Food -- Analysis -- Periodicals; 664.005
• DOI: 10.1111/jfpe.14303
• Languages: English
• ISSNs: 0145-8876

Abstract

Abstract: Different quality grades of tea tend to have a high degree of similarity in appearance. Traditional image‐based identification methods have limited effects, while complex deep learning architectures require much data and long‐term training. In this paper, two tea quality identification methods based on deep convolutional neural networks and transfer learning are proposed. Different types and quality of tea images are collected by a self‐designed computer vision system to form a data set, which is small‐scale and of high inter‐ and intraclass similarity. The first method uses three simplified convolutional neural network (CNN) models with different image input sizes to identify the quality of tea. The second method performs transfer learning to identify the tea quality by fine‐tuning the mature AlexNet and ResNet50 architecture. Classification performance and model complexity are measured and compared. The related application software is also developed. The results show that the performance of the CNN models and the transfer learning models are close, and both can achieve high identification accuracy. However, the complexity of the CNN models is two to three orders of magnitude lower than that of the transfer learning models. The study shows that deep CNNs and transfer learning have great potential to be rapid and effective methods for automated tea quality identification tasks with high inter‐ and intrasimilarity. Abstract :