Abstract: Abstract: To address the challenges of the scarcity of facial diagnosis datasets and the difficulty in learning facial color features, this paper proposes a facial color diagnosis classification network optimized by meta-learning, aiming to enhance the network’s adaptability to small datasets and reduce its reliance on large amounts of training data. Facial diagnosis, as an essential component of Traditional Chinese Medicine (TCM), assists in assessing health conditions by observing changes in facial color. However, existing facial diagnosis data resources are limited, and the sparsity and diversity of features impose higher demands on the models. First, a meta-learning algorithm is applied to initialize the network’s parameters, enabling the network to have a strong learning capability from the outset, and thus more effectively capture key features in the limited facial diagnosis data. The proposed classification network structure incorporates a residual convolutional module, which is pruned during the inference stage using reparameterization techniques, reducing the computational complexity and the number of parameter count. To validate the effectiveness of the proposed method, experiments were conducted on two clinical facial color classification datasets, Face3c and Face5c. A comparative analysis was made with several current mainstream classification methods. The results show that the proposed classification network consistently outperforms other methods across all evaluation metrics, accuracy scores of 93.67% and 81.13%, respectively. This result indicates that the proposed method can effectively address the classification challenges of small facial diagnosis datasets, offering new possibilities for the intelligent development of traditional Chinese medicine diagnostics.

Key words: Key words: meta-learning, facial color classification, deep learning, facial diagnosis