Comparative Analysis of Lightweight Vision Transformers and CNNs for Efficient Bacterial Species Classification

Food safety requires rapid and accurate bacterial identification to prevent disease and economic losses. This study compares three lightweight deep learning models—Tiny-ViT, ShuffleNetV2, and EfficientNet-Lite—for classifying 33 bacterial species from a combined public dataset. Models were trained using transfer learning with original and augmented data and evaluated using 5-fold cross-validation. Tiny-ViT achieved the highest performance with 99.66% accuracy and 99.70% precision, setting a new state-of-the-art for the DIBaS dataset. EfficientNet-Lite reached 99.32% accuracy with superior efficiency—threefold lower FLOPs (397.49M), fewer parameters (3.41M), and faster inference (0.90 ms/image). Comparison of per-class error rates across four models—Tiny-ViT Original, Tiny-ViT Augmented, EfficientNet-Lite Augmented, and EfficientNet-Lite Original—showed consistent stability, where each bacterial class exhibited low mean error and narrow 95% confidence intervals (CI95%), reflecting statistical reliability. These findings highlight a trade-off: Tiny-ViT offers maximum accuracy, while EfficientNet-Lite provides optimal accuracy–efficiency balance for edge-based bacterial diagnostics.