摘要
针对卷积神经网络处理视线预测数据集时,在高迭代次数下产生的过拟合和精准度不足问题,提出面向视线预测的双通道残差网络。首先,将数据集中人脸的左眼与右眼分别用残差网络进行训练;然后,将双通道残差网络计算得到的局部特征通过权值矩阵计算一起作为输入连接到下一级全连接层;最后,经过2个全连接层对特征向量进行分类提取,得到更理想的输出结果。在数据集GazeCapture和MPIIGaze上的训练结果表明:采用双通道网络结构比单通道网络结构进行视线预测的误差小;样本容量大的数据集可有效提高预测精准度;样本筛选后双通道残差网络可在更短的周期内达到稳定拟合状态,并提高预测准确度,具有更好的可靠性和鲁棒性。
In order to solve the problem of over-fitting and insufficient accuracy caused by high iteration times when the convolutional neural network processes the data set of gaze estimation,a double-channel residual network for gaze estimation is proposed.The left eye and right eye of the face in the dataset were trained by residual network respectively,and then the local features obtained by the calculation of the double-channel residual network were connected to the next fully connection layer as input through the calculation of the weight matrix.The feature vectors are classified and extracted through two fully connection layers,and a better output result is obtained.The results of the training on GazeCapture and MPIIGaze show that compared with the single channel network structure,the two channel network structure has smaller error in line of sight prediction;the data set with large sample size can effectively improve the prediction accuracy;after sample screening,the two channel residual network can reach a stable fitting state in a shorter period,and improve the prediction accuracy,with better reliability and robustness.
作者
杨春雨
文元美
Yang Chunyu;Wen Yuanmei(School of Information Engineering,Guangdong University of Technology,Guangzhou 510006,China)
出处
《自动化与信息工程》
2020年第1期10-15,21,共7页
Automation & Information Engineering
基金
国家自然科学基金资助项目（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）
关键词
视线预测
残差网络
非线性回归问题
全连接层
gaze estimation
residual network
nonlinear regression problem
fully connected layer
