High-end Image Nmi Value-based Gesture Recognition Algorithm

The Research of Gesture-Language can be applied in many fields such as Computer aided Gesture-Language Teaching, TV Bilingual Broadcasting, the research of Virtual Human. The research of Gesture-Language includes the following subjects: Education, Computer Graphology, Robot Motion and Physic etc. It is a very meaningful subject. The reearch of Gesture Recognition has a wide range of applications such as: the communication between the deaf and the normal, the aided recognition of voice recognition ,the control of VR, the study of robot.This paper discussed the research of vision-based Gesture Recognition based in 3 aspects: gesture image preprocessing, feature extraction and the design of classifier.In the process of image preprocessing there are several image operations. Firstly, The system is to turn RGB colour images taken from digital camera into gray-scaled images. Then System takes the smoothing measure with Guass template to reduce the noise in the images. After that, It gets the binary version of the images by the means of KSW entropy. Lastly, to get a better binary image the system takes the operation of region growing.Following the image preprocessing, it's turn to extract the right feature from the gesture. The uncertainty of rotation and scale of gesture makes us many difficulties to the extraction of feature. Based on this, system grossly classifies the gesture according to the orientation of the gesture, and then it calculates the high-order NMI value of the gesture image because of its invariability of image rotaton and scale.In the classifier designing, system splits the 9 sets of gesture images into one testing set and 8 designing sets fistly. Then it calssifies the input gesture according to its orientation and calculate the NMI value of the unknown gesture. Finally, comparing with the feature of designing sets, it gets the result by K-nearest neighbor rule with refusal dicision.The wrongly-classified rate is only 6.67% and the refusal rate 10% when our system classify the whole testing set. The result shows our system is totally effective.