A lightweight method for tooth-marked tongue recognition based on convolutional neural networks

The tooth-marked tongue is an important objective indicator in Traditional Chinese Medicine (TCM) tongue diagnosis, and its automatic recognition provides a computationally inexpensive route towards instrumented tongue inspection. To address the problem that existing tooth-mark recognition methods have large model sizes and high computational cost, this paper proposes a lightweight tooth-marked tongue recognition method. The method performs binary tooth-mark classification on pre-segmented tongue images: a lightweight convolutional neural network, MobileNetV3-Small, is combined with transfer learning and data augmentation to perform binary tooth-mark classification, while a class-weighted cross-entropy loss is introduced to alleviate class imbalance. Experiments on a public tooth-marked tongue dataset (1250 images), evaluated under a stratified 70:15:15 partition in which the test set is held out from model selection and averaged over 10 repetitions, show that the proposed method achieves an accuracy of 85.32 ± 1.49% and an F1 of 84.89 ± 1.47%, with only 1.52M parameters (about one seventh of ResNet18), 0.061 GFLOPs, and a model size of 6.2 MB, substantially reducing model complexity while maintaining comparable accuracy. Ablation studies identify transfer learning as the decisive component, and Grad-CAM visualization further shows that the model focuses on the tongue edge where tooth marks appear, indicating good interpretability.