Research On Segmentation And Feature Extraction Method For Visual Language Recognition

With the rapid development of human-computer interaction technology,a variety of intelligent devices come into our life gradually in recent years.Furthermore,the artificial intelligence has become the research focus in information field formally.As a proven human-computer interaction technology,speech recognition can benefit our live tremendously.However,in a noisy environment,the recognition rate of automatic speech recognition is drop dramatically,and this system cannot complete the operation.Visual lip movement also has discourse content information in the conversation.Thus,researchers commence commit to researching the discourse content recognition based on lip vision which is called lip moving technique research,it plays an important role in remedy a defect of automatic speech recognition.This paper focuses on lip segmentation and feature extraction in lip reading systems.For the lip segmentation,the OpenCV detection technique is used to detect and locate faces in the image firstly,the lip bounding box is located and is transformed to different color spaces which include RGB,HSV and chromatic.Secondly,color enhancement,color segmentation and lip corners location are performed in this three color spaces over the lip image.Finally,the method of BFOA and Kapur maximum entropy threshold method are applied to obtain the best threshold of the lip region,in order to achieve accurate segmentation of lip.For the visual feature extraction,pixel characteristics lose their original application value due to the difference of skin color among the races and the different light conditions.In this paper,the key points on the lip contour are located based on accurate segmentation of lip.Using these key points as inputs,we fit the outline of lip contour based on least square criterion and extract desired lip feature vector precisely.At the same time,the convex hull algorithm is applied to the lip outline in order to get accurate results.Experimental result proves that the proposed method has strong robustness and higher accuracy,improved the accuracy of features extraction simultaneously.