Comparison of Efficient Deep Learning Architectures for Lactobacillus Species Identification

Identifying microorganism species, such as Lactobacillus, is essential in ensuring the food products' quality and safety. Traditional laboratory practice requires expert knowledge and experience, but the method is expensive and time-consuming due to complex sample preparation. Faster, more accurate, and cheaper computational methods, such as transfer learning technology, are needed for the Lactobacillus species classification. The technique has been effective in a variety of image recognition contexts. Deep learning architecture can also be applied as an innovative strategy for digital image-based identification. Therefore, this research aims to compare several deep-learning architectures in classifying bacterial strains of Lactobacillus. The four architectures (Inception V3, MNASNet, RegNet, and Xception) used have excellent performance with accuracy above 95%. Among those architectures, the mobile neural architecture search network (MNASNet) exhibits the most potential, with 99.15% accuracy, 99.09% precision, 99.14% sensitivity, and 99.11% F1 score. This performance is particularly notable given that MNASNet operates with 3.1 million parameters. Although RegNet has a slightly lower parameter count at 2.68 million, it achieves the accuracy, precision, sensitivity, and F1 score of 98.73%, 98.73%, 98.77%, and 98.75%, respectively. Xception, with over 20 million parameters, attains 98.99% accuracy, 98.71% precision, 98.78% sensitivity, and 98.74% F1 score. In addition, the MNASNet architecture is highly efficient, making it very suitable to be implemented or embedded on mobile devices. The model is promising regarding the deep learning architecture prospects for classifying microscopic images of Lactobacillus species.