Lightweight Gaze Estimation Method Based on Multi-level Feature Refining
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
维普
万方数据
近年来,基于外观的注视估计取得了显著进展.然而,现有方法多以庞大的网络参数量为代价来提高精度,使得模型的开发部署成本昂贵.针对此问题,提出一种基于多级特征提炼的轻量化注视估计网络(Lightweight Network with Multi-level Feature Refining,LMLFR-Net).其包含一种轻量级特征提取模块(SECA)和一种轻量级多级特征提炼模块(FRM).SECA融合了挤压激励和协调注意力,以提高模型对特征的精化能力;FRM将主干网络的多级特征进行融合提炼,通过同时利用低层与高层特征,提升了模型对细节的捕获能力,在不显著增加参数量的同时,改善轻量级网络的估计精度.实验表明,所提出的网络在MPIIFaceGaze数据集上的估计精度相比FAR-Net提升了2.14%,参数量减少了85.35%,表现出了良好的轻量化性能.
Appearance-based gaze estimation has made significant progress in recent years.However,existing methods mostly improve accuracy at the expense of a huge amount of network parameters.This makes the development and deployment cost of the model expensive.In view of this problem,a lightweight gaze estimation network based on multi-level feature refining(Lightweight Network with Multi-level Feature Refining,LMLFR-Net)is proposed.It includes a lightweight feature extraction module(SECA)and a lightweight multi-level Feature Refining Module(FRM).SECA combines Squeeze-and-Excitation and Coordinate Attention to improve the model's ability to refine features.FRM integrates and refines the multi-level features of the backbone network,and improves the model's ability to capture details by simultaneously utilizing low-level and high-level features.It improves the estimation accuracy of lightweight networks without significantly increasing the number of parameters.Experiments show that the estimation accuracy of the proposed network on the MPIIFaceGaze data set is improved by 2.14%compared to FAR-Net,and the number of parameters is reduced by 85.35%,showing good lightweight performance.
维普
万方数据
