Lip Segmentation for Visual Speech Recognition Based on the Convolution Process

Lip segmentation is an important process in visual speech recognition. Nevertheless, in many cases, the poor isolation limited the recognition rate. Therefore, this paper proposes a new isolation method based on the mean and the convolution technique with a special kernel to enhance the isolation process. The database that is used in this paper is the UOTletters, which consists of 30 speakers, 26 English letters (repeated twice for each speaker), and the overall utterances are 1560. The experiment results have shown that the proposed algorithm has higher accuracy (100^%) with respect to the related works. Consequently, this algorithm is recommended for geometric feature extraction in visual speech recognition due to its higher accuracy.