Research On Lip Recognition Technology

Lip-reading technology can be the speaker information capture through the lip motion,the technology has been widely applied in the field of speech recognition, identification,human-computer interface and the multimedia system. A complete lip-reading systemmainly consists of three units: lip detecting and locating the lip, lip reading recognition,feature extraction. This paper mainly lip to lip reading system detection unit and lip featureextraction unit was studied in detail.Detecting and locating the lip is one of the most important links in lip-readingrecognition. Because of the lip color features, extensive research carried out at home andabroad by detecting the lip color characteristics, and achieved certain results, but there is noperfect lip detection algorithm. In this paper, color image based on face, color distributionon the different races of the lip color and skin color was studied in detail, puts forward anadaptive filtering algorithm based on color information, according to the geometricalcharacteristics and lip in the face of the lip in the YCrCb space color separationcharacteristics of the color filter, adaptive, effective segmentation the lip contour, accuratedetection and extract the lip contour. Results validate the algorithm, two groups ofexperiments, comparing the proposed algorithm and used for lip detection Red Exclusionalgorithm. The experimental results show that, the method in this paper has the obviousimprovement in efficiency, robustness and the support of different human complexion, andthe algorithm complexity is not increased significantly.Feature extraction plays an important role in lip-reading, goal is to obtain a lowdimensional feature vector, low redundancy and representative. Pixel based featureextraction methods are studied in this paper, the characteristics of a cascade of extractionprocess, the corresponding transformation of the image, and then to transform the results ofdimension reduction, and feature normalization. Comparison and analysis of severaltransformation method based on the PCA, DCT-PCA and Gabor-PCA dimension of DCTand Gabor wavelet transform, optimal recognition rate can reach77.4%and77.9%respectively, compared with the direct selection method of transform coefficient increasesthe recognition rate of about10%.