基于多任务卷积神经网络的实时人眼检测方法

Real-Time Eye Detection Based on Multi-Task Convolutional Neural Networks

针对戴眼镜、人脸姿态变化以及眯眼睛等复杂场景,提出了一种基于多任务卷积神经网络(MultiTask CascadedConvolutionalNetworks,MTCNN)的人眼检测算法。针对性地调整与优化网络,删除landmark部分以简化网络结构,进而调整网络的输入尺寸,使模型更适用于人眼检测。实验结果表明,基于MTCNN的人眼检测算法在数据集上准确率达92.1%,图形处理器(GraphicsProcessingUnit,GPU)检测速度达112frames/s,可以有效兼顾实时性与准确性。

For complex scenes such as wearing glasses,face pose changes and squinting,this paper proposes a human eye detection algorithm based on MTCNN(Multi Task Cascaded Convolutional Networks).In this paper,the network is adjusted and optimized,the landmark part is deleted to simplify the network structure,the input size of the network is adjusted to make the model more suitable for human eye detection,and the structure of the model is adjusted to improve the performance of human eye detection.The experimental results show that the accuracy of the algorithm proposed in this paper is 92.1%on data set,and the Graphics Processing Unit(GPU)detection speed is 112 frames/s.It can effectively give consideration to real-time and accuracy.