Abstract:Accurate gaze direction estimation is a crucial technology in various applications such as human-computer interaction and virtual reality. The appearance-based gaze estimation is currently the mainstream method. However，due to the diversity in eye appearance，lighting conditions，and head poses，gaze estimation in unconstrained environments remains a challenging task.This paper proposes a high-frequency amplifier gaze estimation network（HFA-Net）.Initially incorporating a high-frequency information extraction module and Convolutional Block Attention Module（CBAM）into the neural network helps reduce the impact of redundant information. Finally，the network divides gaze angles into two separate regression tasks，each optimized with independent loss functions. The proposed method is trained and tested on the publicly available MPIIGaze dataset，and experimental results demonstrate that this approach achieves the best angle estimation error of 4.17° on MPIIGaze，surpassing the current mainstream algorithms.