Facial expressions are a prevalent way to recognize human emotions, and automatic facial expression recognition (FER) has been a significant task in cognitive science, artificial intelligence, and computer vision. The critical issue with the design of the FER model is the strong intra-class correlation of different emotions. The accuracy of the FER model is reduced due to other problems such as the variations in expressing the emotions, variations in lighting, and different ethnic biases. The latest convolutional neural network-based FER models have shown significant improvement in accuracy score but lack distinguishing the micro-expressions. This paper proposed a multi-input hybrid FER model that considers both hand-engineered and self-learnt features to classify facial expressions. The VGG-Face and the histogram of oriented gradients (HOG) features are derived from the faces to distinguish various facial expression patterns. The fusion of deep (VGG-Face) and hand-engineered (HOG) features has shown improved accuracy compared to the conventional CNN models. The results obtained showed that the proposed model’s accuracy scores outperformed the accuracy scores of the other popular FER models on three facial expression datasets. Extended Cohn–Kanade (CK\(+\)), Yale-Face, and Karolinska directed emotional faces (KDEF) datasets are used to determine the model’s classification efficiency. The proposed model scored 98.12%, 95.26%, and 96.36% accuracy using a fivefold cross-validation process on the CK\(+\), Yale-Face and KDEF datasets.

面部表情是识别人类情绪的一种流行方式，自动面部表情识别（FER）一直是认知科学、人工智能和计算机视觉领域的一项重要任务。FER 模型设计的关键问题是不同情绪的强类内相关性。由于其他问题，例如表达情绪的变化、照明的变化和不同的种族偏见，FER 模型的准确性降低了。最新的基于卷积神经网络的 FER 模型显示出准确度得分的显着提高，但缺乏区分微表情。本文提出了一种多输入混合 FER 模型，该模型同时考虑了手工设计和自学习的特征来对面部表情进行分类。VGG-Face和定向梯度直方图（HOG）特征来源于人脸，以区分各种面部表情模式。与传统的 CNN 模型相比，深度（VGG-Face）和手工工程（HOG）特征的融合显示出更高的准确性。获得的结果表明，该模型在三个面部表情数据集上的准确度得分优于其他流行的 FER 模型的准确度得分。扩展的 Cohn-Kanade (CK 获得的结果表明，该模型在三个面部表情数据集上的准确度得分优于其他流行的 FER 模型的准确度得分。扩展的 Cohn-Kanade (CK 获得的结果表明，该模型在三个面部表情数据集上的准确度得分优于其他流行的 FER 模型的准确度得分。扩展的 Cohn-Kanade (CK\(+\) )、Yale-Face 和 Karolinska 定向情绪面孔 (KDEF) 数据集用于确定模型的分类效率。所提出的模型在 CK \(+\)、Yale-Face 和 KDEF 数据集 上使用五重交叉验证过程获得了 98.12%、95.26% 和 96.36% 的准确度。