A novel lip localization method based on shiftable wavelets transform

Visual feature extraction is one of the most important techniques in audiovisual bimodal speech recognition, and also remains a very challenging area in image understanding. A shiftable multiscale transform is introduced into the construction of an active shape model. It uses the pyramidal data to describe the structure of an image, which is invariant to illumination and perspective variability and thus contributes a lot to the improvement of the robustness of the model. A segmental downhill simplex method is also put forward to improve the minimization procedure of lip localization. It employs a kind of “coarse-to-fine” strategy to speed up the convergence and improve the robustness of lip localization. Experiments support the validity of the new method, and show better robustness and higher efficiency