Research on a Lip Reading Algorithm Based on Efficient-GhostNet

Lip reading technology refers to the analysis of visual information speaker’s mouth movements recognize content speech. As one important aspects human–computer interaction, lip has gradually become popular with development deep learning in recent years. At present, most networks are very complex, large numbers parameters and computation, model generated by training needs occupy memory, which brings difficulties for devices limited storage capacity computation power, such as mobile terminals. Based on above problems, this paper optimizes improves GhostNet, a lightweight network, it proposing more efficient Efficient-GhostNet, achieves performance improvement while reducing number through local cross-channel interaction strategy, without dimensionality reduction. The improved Efficient-GhostNet is used perform spatial feature extraction, then extracted features inputted GRU network obtain temporal sequences, finally prediction. We Asian volunteers recording dataset paper, also adopting data enhancement dataset, using angle transformation deflect process recorder 15 degrees each left right, order be able enhance robustness better reduce influence other factors, well improve generalization ability so that can consistent recognition scenarios real life. Experiments prove + achieve purpose comparable accuracy.