Class-balanced and Local Median Loss Jointly Supervised for Wild Facial Expression Recognition
首发时间:2020-02-14
Abstract:Over the past few years, Convolutional Neural Networks (CNNs) have shown effective performance on facial expression recognition. However, it is still a challenge problem for facial expression in the wild. The facial expression recognition dataset in the wild usually has the problem that imbalanced distribution of facial expression data and large intra-class differences caused by factors such as pose, lighting and gender. In order to solve this problem, this paper presents a novel loss function -- CALM Loss (Class-balanced and Local Median Loss). CALM Loss contains two parts. The first part is the class-balanced softmax loss function, which is uesd to solve the problem of data imbalance. The data set is divided into two classes, with two expressions with less data as one class and the other five as one class. During the network training process, the weight of the class with less data is adaptively increased. The second part is the local median loss function, which uses the median of serveral neatest neighbors in the same class as the center of class, weakens the influence of difficult samples on the selection of class center. Finally, the CALM Loss training network was adopted in this paper. The average recognition accuracy on the RAF dataset reaches 77.34$\%$, which proves the effectiveness of the proposed method.
keywords: facial expression recognition imbalanced dataset intra-class variantion loss function CNN
点击查看论文中文信息
类别均衡与局部中值损失联合监督的自然场景中人脸表情识别
摘要:近年来,卷积神经网络在实验室控制环境下的人脸表情识别任务中取得了很大进步,但是在自然场景中人脸表情识别仍然存在一些挑战.针对自然场景中人脸表情数据分布不平衡,以及由姿势、光照和性别等因素引起的类内差异大的问题,本文提出了一个新的损失函数 CALM Loss (Class-balanced and Local Median Loss).CALM Loss包含两个部分,第一部分是类别均衡softmax损失函数,用来解决数据分布不平衡的问题.将数据集分为两大类,数据量少的两种表情作为一类,其余五种作为一类,在网络训练过程中自适应地增大数据量少的类别的权重.第二部分是局部中值损失函数,采用同一个类别内每个样本的若干近邻的中值作为该类别的中心,减弱难样本对类别中心选择的影响,从而使类内样本更紧凑.最后,本文采用CALMLoss训练网络并得到分类结果,将RAF数据集上的平均识别准确率提高到了77.34$\%$,证明了本文方法的有效性.
引用
No.****
动态公开评议
共计0人参与
勘误表
类别均衡与局部中值损失联合监督的自然场景中人脸表情识别
评论
全部评论