Gaze Estimation Based on the Improved Xception Network

The estimation of gaze points plays a crucial role in various fields, from scientific research to commercial product applications, aiding in the understanding of human behavior and enhancing human–machine interactions. Research on appearance-based gaze estimation has shown that convolutional neural networks (CNNs) can effectively extract image features, simplifying the gaze estimation process. Using facial or ocular images as input, we can train a mapping model between appearance and gaze to determine the corresponding gaze points. This model aims to improve classification accuracy and enhance model stability. To achieve this goal, we introduce attention mechanisms separately, optimize the model to improve its object localization capabilities and increase accuracy, and analyze the influence of each attention mechanism on the network model. An automatic network structure search is used to optimize the model pruning method, aiming to improve the model’s inference speed and reduce storage space while maintaining accuracy. The fast gradient sign method (FGSM) algorithm is used to test the model’s stability. The improved Xception network outperforms the original network in various metrics, reducing its hardware requirements and enabling deep CNNs to be used on mobile and embedded devices. This advancement holds significant value for practical human–machine interaction applications.