基于卷积神经网络和池化算法的表情识别研究  

Research on Emotional Recognition Based on Convolution Neural Network and Pooling Algorithm

作　　者：陈航 邱晓晖[1] CHEN Hang;QIU Xiao-hui(School of Telecommunications & Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003

出　　处：《计算机技术与发展》2019年第1期61-65,共5页Computer Technology and Development

基　　金：江苏省自然科学基金(BK2011789);东南大学毫米波国家重点实验室开放课题(K201318)

摘　　要：卷积神经网络(CNN)能够通过神经网络自主学习提取图像中的特征,并且具有局部响应、权值共享等优点,在人脸表情识别中获得了广泛的应用。池化算法是CNN的核心技术之一,通过对卷积层的特征进行聚合统计,池化算法可以减少CNN的特征维度,提高特征表征能力,但是目前常用的池化算法还存在提取特征单一,缺乏灵活性的情况。为了克服现有池化算法的不足,根据深度学习可采用BP算法自主调节参数的特性,提出一种改进的自适应池化算法。该算法在训练过程中能够根据损失函数,不断更新池化域的参数,最终使表情预测值和真实结果值之间的差值达到最小。基于CK+人脸表情数据库的实验结果表明,与现有池化算法相比,提出的自适应池化算法能有效提高表情识别准确率。Convolution neural network(CNN)can extract the features of the image by neural network autonomy learning with the advantages of local response and weight sharing,and has been widely used in facial expression recognition.The pooling algorithm is one of the core technologies of CNN.By aggregating the features of the convolution layer,the pooling algorithm can reduce the feature dimension of the CNN and improve the capability of feature representation.However,the commonly used pooling algorithms also have the disadvantage of single extraction feature and lack of flexibility.For this,we propose a modified adaptive pooling algorithm based on the characteristics of depth self-tuning parameters which can be adjusted by BP algorithm.According to the loss function,updating the parameters of the pooled domain eventually minimizes the difference between the predictive value of the expression and the true result value.The experiment based on CK+face expression database shows that compared with the existing pooling algorithm,the proposed adaptive pooling algorithm can effectively improve the accuracy of facial expression recognition.

关 键 词：卷积神经网络 池化算法 人脸表情识别 深度学习 特征提取 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]