Inner lips feature extraction based on CLNF with hybrid dynamic template for Cued Speech

摘要: In previous French Cued Speech (CS) studies, one of the widely used methods is painting blue color on speaker’s lips to make feature extraction easier. this paper, in order get rid artifice, a novel automatic method extract inner contour CS speakers presented. This based recent facial model developed computer vision, called Constrained Local Neural Field (CLNF), which provides eight characteristic landmarks describing contour. However, directly applied our data, CLNF fails about 41.4% cases. Therefore, we propose two correct B parameter (aperture lips) and A (width lips), respectively. For correcting parameter, hybrid dynamic correlation template (HD-CTM) using first derivative smoothed luminance variation proposed. HD-CTM detect outer lower position. Then, position obtained by subtracting validated thickness (VLLT). periodical spline interpolation with geometrical deformation six explored. Combined an round detector, efficient for (the third vowel viseme made vowels small opening). evaluated 4800 images three speakers. It corrects 95% errors total RMSE pixel (i.e., 0.05 cm average) achieved. The tested 927 images. error reduced significantly, comparable state art. Moreover, properly distributed plane after method.