Cao Xiyuan, Zhang Delong, Jin Chunyang, Zhang Zhidong, Xue Chenyang
State Key Laboratory of Extreme Environment Optoelectronic Dynamic Testing Technology and Instrument, North University of China, Taiyuan 030051, China.
Biomimetics (Basel). 2025 Jun 13;10(6):402. doi: 10.3390/biomimetics10060402.
Variations in facial complexion serve as a telltale sign of underlying health conditions. Precisely categorizing facial complexions poses a significant challenge due to the subtle distinctions in facial features. Three multi-feature facial complexion classification algorithms leveraging convolutional neural networks (CNNs) are proposed. They fuse, splice, or independently train the features extracted from distinct facial regions of interest (ROI), respectively. Innovative frameworks of the three algorithms can more effectively exploit facial features, improving the utilization rate of feature information and classification performance. We trained and validated the three algorithms on the dataset consisting of 721 facial images that we had collected and preprocessed. The comprehensive evaluation reveals that multi-feature fusion and splicing classification algorithms achieve accuracies of 95.98% and 93.76%, respectively. The optimal approach combining multi-feature CNN with machine learning algorithms attains a remarkable accuracy of 97.78%. Additionally, these experiments proved that the multidomain combination was crucial, and the arrangement of ROI features, including the nose, forehead, philtrum, and right and left cheek, was the optimal choice for classification. Furthermore, we employed the EfficientNet model for training on the face image as a whole, which achieves a classification accuracy of 89.37%. The difference in accuracy underscores the superiority and efficacy of multi-feature classification algorithms. The employment of multi-feature fusion algorithms in facial complexion classification holds substantial advantages, ushering in fresh research directions in the field of facial complexion classification and deep learning.
面部肤色的变化是潜在健康状况的明显迹象。由于面部特征的细微差别，精确对面部肤色进行分类面临重大挑战。提出了三种利用卷积神经网络（CNN）的多特征面部肤色分类算法。它们分别融合、拼接或独立训练从不同面部感兴趣区域（ROI）提取的特征。这三种算法的创新框架能够更有效地利用面部特征，提高特征信息的利用率和分类性能。我们在由我们收集并预处理的721张面部图像组成的数据集上对这三种算法进行了训练和验证。综合评估表明，多特征融合和拼接分类算法的准确率分别达到95.98%和93.76%。将多特征CNN与机器学习算法相结合的最优方法达到了97.78%的显著准确率。此外，这些实验证明多域组合至关重要，包括鼻子、额头、人中以及左右脸颊在内的ROI特征排列是分类的最佳选择。此外，我们使用EfficientNet模型对整个面部图像进行训练，其分类准确率为89.37%。准确率的差异凸显了多特征分类算法的优越性和有效性。多特征融合算法在面部肤色分类中的应用具有显著优势，为面部肤色分类和深度学习领域带来了新的研究方向。
