A Residual Dense Lightweight Group Convolution Neural Network for Identifying the Gas Information of Different Levels of Tea

In the tea market, it is common for low-quality tea to cheat consumers by pretending to be high-quality. This work proposes a fast and nondestructive tea quality detection method, which combines an electronic nose (e-nose) system with an adaptive gas information recognition method. First, based on a PEN3 e-nose system, the gas information of different levels of tea is obtained. Second, a lightweight group convolution (LGC) module is proposed to adaptively focus on the important features that affect the classification performance of gas information and reduce the number of parameters. Finally, residual dense block (RDB) is introduced to fuse the shallow and deep features to avoid feature degradation, and residual LGC neural network (RLGCNet) is designed to recognize the different levels of tea gas information effectively. In the comparison results of the ablation study and multiclassification model, RLGCNet obtained the best classification accuracy of 98.50%, the precision of 98.49%, and the recall of 97.69%. In conclusion, the theoretical research results provide an effective detection method for the quality supervision of the tea market.