EfficientNet-Gaze: Integrating Multi-Scale Feature Extraction with Frequency Domain Analysis for Efficient Gaze Estimation

Abstract

In recent years, deep learning-based appearance-based gaze estimation methods have gained increasing popularity because of their high accuracy and fewer environmental constraints. However, existing high-accuracy models typically rely on deeper network structures and mainly focus on processing in the spatial domain, resulting in a large number of parameters and high computational costs, which limit their practical application in resource-constrained environments. Therefore, the paper proposes a novel lightweight model, EfficientNet-Gaze, which significantly reduces computational complexity while achieving accurate estimation of gaze angle. EfficientNet-Gaze combines multi-scale feature extraction and frequency domain analysis techniques, utilizing dilated convolutions and depthwise convolutions to enhance feature extraction capabilities. It also employs Fast Fourier Transform (FFT) to process feature maps in the frequency domain, which not only improves computational efficiency but also enhances the model’s ability to extract global features. To validate the performance of the model, extensive experiments are conducted on several public datasets. The results show that, compared with the latest gaze estimation method Gaze-Swin, EfficientNet-Gaze achieves a lower angular error (3.87 ° in MPIIFaceGaze, 6.23 ° in RT-Gene, and 10.46 ° in Gaze360) with a reduction of 94. 02% and 92. 45%, respectively. These findings indicate that EfficientNet-Gaze has broad application potential in gaze estimation tasks.