Siddiqi Muhammad Hameed, Ahmad Irshad, Alhwaiti Yousef, Khan Faheem
IEEE J Biomed Health Inform. 2025 Jan;29(1):30-42. doi: 10.1109/JBHI.2024.3482450. Epub 2025 Jan 7.
Facial expressions vary with different health conditions, making a facial expression recognition (FER) system valuable within a healthcare framework. Achieving accurate recognition of facial expressions is a considerable challenge due to the difficulty in capturing subtle features. This research introduced an ensemble neural random forest method that utilizes convolutional neural network (CNN) architecture for feature extraction and optimized random forest for classification. For feature extraction, four convolutional layers with different numbers of filters and kernel sizes are used. Further, the maxpooling, batch normalization, and dropout layers are used in the model to expedite the process of feature extraction and avoid the overfitting of the model. The extracted features are provided to the optimized random forest for classification, which is based on the number of trees, criterion, maximum tree depth, maximum terminal nodes, minimum sample split, and maximum features per tree, and applied to the classification process. To demonstrate the significance of the proposed model, we conducted a thorough assessment of the proposed neural random forest through an extensive experiment encompassing six publicly available datasets. The remarkable weighted average recognition rate of 97.3% achieved across these diverse datasets highlights the effectiveness of our approach in the context of FER systems.
面部表情会因不同的健康状况而有所变化，这使得面部表情识别（FER）系统在医疗保健框架中具有重要价值。由于难以捕捉细微特征，实现对面部表情的准确识别是一项相当大的挑战。本研究引入了一种集成神经随机森林方法，该方法利用卷积神经网络（CNN）架构进行特征提取，并使用优化的随机森林进行分类。在特征提取方面，使用了具有不同滤波器数量和内核大小的四个卷积层。此外，模型中还使用了最大池化、批量归一化和随机失活层，以加快特征提取过程并避免模型过拟合。提取的特征被提供给优化的随机森林进行分类，该随机森林基于树的数量、准则、最大树深度、最大终端节点、最小样本分割和每棵树的最大特征数，并应用于分类过程。为了证明所提出模型的重要性，我们通过涵盖六个公开可用数据集的广泛实验，对所提出的神经随机森林进行了全面评估。在这些不同数据集上实现的高达97.3%的显著加权平均识别率，突出了我们的方法在FER系统背景下的有效性。
