一种面向表情识别的ROI区域二级投票机制  

Expression-oriented ROI region secondary voting mechanism

作　　者：文元美[1] 欧阳文 凌永权 Wen Yuanmei;Ouyang Wen;Ling Yongquan(School of Information Engineering,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学信息工程学院

出　　处：《计算机应用研究》2019年第9期2861-2865,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61671163,11871168);广东省自然科学基金项目(2018A030310593);2017年中央财政支持地方高校发展专项资金项目(201707)

摘　　要：针对如何更有效地使用卷积神经网络从训练图像中学习到的分布式特征进行研究,提出了一种面向人脸表情识别的ROI区域二级投票机制。将图像划分成一系列感兴趣区域(ROI)图像输入到卷积神经网络中进行训练;然后将测试图像的ROI图像输入到卷积神经网络中,统计所有ROI图像的判别结果;最后采用二级投票机制确定测试图像的最终类别,得到最终判别结果。针对卷积神经网络不能从人脸图像中学习到旋转等空间位置信息,引入了STN(spatial transformer network),提高算法在解决复杂情况下的表情识别问题的能力。实验表明,ROI区域二级投票机制能够更有效地使用卷积神经网络从训练图像中学习到的分布式特征,比直接使用ROI图像进行投票的方法准确率提升了1. 1%,引入STN能够有效提升卷积神经网络的鲁棒性,比未引入STN的方法准确率提升了1. 5%。Aiming at the problem of how to more efficiently use the distributed features that convolutional neural network has learned from training images,this paper proposed a regions of interest( ROI) region secondary voting mechanism for facial expression recognition. Firstly,it divided the image into a series of ROI images,which were input into the convolutional neural network for training. Then,it input the ROI images of the test image into the convolutional neural network and got all ROI images’ discriminant results. Lastly,it used the secondary voting mechanism to determine the final category of test image. In addition,aiming at the problem of convolutional neural network cannot learn spatial position information such as rotation,this paper introduced the spatial transformer network to make convolutional neural network useful in complex condition. Experiments show the ROI region secondary voting mechanism can more effectively use the distributed features which learned by convolutional neural network,compared with the method of voting directly using ROI images,the accuracy is increased by 1. 1%. The introduction of STN can effectively improve the robustness of convolutional neural network,compared with non-introduced STN,the accuracy is increased by 1. 5%.

关 键 词：卷积神经网络 表情识别 空间变换网络 二级投票机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]