Human facial expression is the core carrier of feedback. Facial expression recognition(FER) has been introduced into mickle fields, such as auxiliary medical care, safe driving, marketing assistance, distance education. However, in the real production process, facial expression image samples collected in different scenarios have problems such as complex backgrounds, which causes the FER model to train very slowly, low recognition rate, and insufficient generalization, so it cannot meet the actual production requirements. As the originator of the clustering algorithm, fuzzy C-means clustering(FCM) algorithm has stable performance and good results. It is applied to the convolutional layer of a convolutional neural network(CNN) to obtain a convolution kernel with an initial value, so as to extract the expression image features in the training set and the test set. This can solve the problem of random initialization of the convolution kernel. Based on the CNN, this paper introduces FCM to optimize the feature extraction (FE) capability of the model, and proposes a novel FER algorithm using an improved CNN(F-CNN). Because traditional CNN has problems such as irrational layer settings and too many parameters. The proposed F-CNN first adjusts the CNN network structure to improve the nonlinear expression ability of CNN. Then, replace the Softmax classifier that comes with CNN with a support vector machine (SVM) to improve the model’s classification ability. The comparison experiments with other models show that the improved model improve the FER rate. The introduced FCM algorithm can effectively improve the model’s FE performance and shorten the time of F-CNN during training. On the whole, F-CNN has reference value.

中文翻译：

---

一种基于改进卷积神经网络的新型面部表情智能识别方法

人的面部表情是反馈的核心载体。面部表情识别（FER）已被引入到医疗辅助、安全驾驶、营销辅助、远程教育等领域。但是，在实际生产过程中，不同场景下采集的人脸表情图像样本存在背景复杂等问题，导致FER模型训练速度非常慢、识别率低、泛化能力不足，无法满足实际生产需求。作为聚类算法的鼻祖，模糊C均值聚类(FCM)算法性能稳定，效果好。将其应用于卷积神经网络（CNN）的卷积层以获得具有初始值的卷积核，从而提取训练集和测试集中的表情图像特征。这样可以解决卷积核随机初始化的问题。本文基于CNN，引入FCM来优化模型的特征提取（FE）能力，并提出了一种使用改进CNN（F-CNN）的新型FER算法。因为传统的CNN存在层设置不合理、参数太多等问题。提出的F-CNN首先对CNN网络结构进行了调整，以提高CNN的非线性表达能力。然后，用支持向量机（SVM）代替CNN自带的Softmax分类器，提高模型的分类能力。与其他模型的对比实验表明，改进后的模型提高了FER率。引入的FCM算法可以有效提高模型的有限元性能，缩短F-CNN在训练过程中的时间。综合来看，F-CNN具有参考价值。