| 基于ShuffleNetV2算法的三维视线估计 |
投稿时间：2021-09-14 修订日期：2021-11-09  点此下载全文 |
| 引用本文： |
| 摘要点击次数: 134 |
| 全文下载次数: 0 |
|
| 基金项目:国家自然科学基金(61663005)。 |
|
| 中文摘要:对近几年来基于神经网络的视线估计方法进行研究, 为了解决当前视线估计网络复杂度较深、精度不高的问题,同时为了未来将网络布署在移动设备端,提出了一种基于ShuffleNet V2算法的视线估计网络,由脸部和眼睛两个子网络构成,脸部子网络通过ResNet v2网络对脸部图片进行特征处理,并加入人脸对齐算法,减少头部角度误差的影响。眼睛子网络通过ShuffleNet V2与ResNet V2算法进行眼睛图片的并行特征处理。网络对特征图片处理后得到角度参数,最后通过坐标变换得到视线角度。并在MPIIGaze数据集上进行了实验。针对精度的不足对算法进行改进,在ShuffleNet V2中加入注意力机制(逐点平方操作模块),并进行了改进算法的验证实验,最后和多种先进的算法进行了实验对比。实验表明,改进后的算法比其他算法的精度要高。 |
| 中文关键词:神经网络 三维视线估计 ShuffleNetV2 ResNet V2 坐标变换 人脸对齐 注意力机制 MPIIGaze |
| |
| 3D gaze estimation based on ShuffleNetV2 algorithm |
|
|
| Abstract:In order to solve the problems of deep complexity and low accuracy of the current line of sight estimation network and to deploy the network on mobile devices in the future, we propose a line of sight estimation network based on ShuffleNet V2 algorithm, which consists of two sub-networks, face and eye, and the face sub-network is used by ResNet v2 network. feature processing of face images and adding face alignment algorithm to reduce the effect of head angle error. The eye sub-network performs parallel feature processing of the eye images through ShuffleNet V2 and ResNet V2 algorithms. The network processes the feature images to get the angle parameters, and finally the angle of view is obtained by coordinate transformation. And experiments were conducted on the MPIIGaze dataset. The algorithm is improved for the lack of accuracy, and the attention mechanism (point-by-point square operation module) is added to ShuffleNet V2, and the verification experiments of the improved algorithm are conducted, and finally the experiments are compared with various advanced algorithms. The experiments show that the improved algorithm has higher accuracy than other algorithms. |
| keywords:Neural networks 3D gaze estimation Shufflenetv2 Resnet V2 Coordinate transformation Face alignment Attention mechanism Mpiigaze |
| 查看全文 查看/发表评论 下载pdf阅读器 |
|
|
|
位访问者,今日访问
次