Sign Language Recognition Method Based On Variable Target Task

Sign language is a way of communication between people. With the development of computer vision technology, sign language gradually become a way of adapting to the social requirement of human computer interaction. Sign language recognition method based on vision has become the important means of human-computer interactive information because of its advantages of intuitive and natural. How to get the accurately classification of the sign language became the scholars’ research focus in recent years. The first thing to realize sign language recognition is to segment gestures from data samples, and then extract gesture features through the effective means under the different algorithms. This paper introduces the key technology of sign language recognition, and do key research on the sign language feature extraction and recognition method in order to improve the gesture recognition accuracy.In this paper, we implement the sign language recognition algorithm based on convolution neural network. Sign language feature extraction is the key to realize recognition tasks. The stand or fall of the extracted features directly influence the outcome of sign language recognition. Convolution neural network is a deep learning algorithms, it has a strong ability to learn. It folds and pools data many times though the network and abstract the features to the identification and classification. But it has much of parameters, large amount of calculation, the ability of obtaining efficient features is limited in the case of limited data samples. This paper improved the optimization, feature extraction and classification method of convolution of the neural network. We improve the efficiency of network optimization and sign language recognition rate with the method that initialization of the network through the existing model and increase the diversity of data samples. And we have has carried on the experiment and analysis of the method.This paper presents a new algorithm for gesture recognition. First we complicated the task though break up the objective function which equivalents to add a regularization item to make the task more challenge. Then training convolution neural network to complete the complex tasks get the data of high performance characteristics. Finally, we train the classifier with the high performance characteristics to the original simple task. We verify the proposed approach based on experiments of the ASL and BSL gesture database, and compare with the gesture recognition method witch already exists. The experimental results show that the method is easy to implement and can improve the efficiency of recognition without any increase in the amount of calculation.