Sign Language Recognition Based On Improved Convolutional Neural Network

Sign language is a medium for deaf people to communicate with the outside world.There are fewer healthy people who have the knowledge of sign language,which inconveniences the communication.Sign language recognition research is a human-computer interaction platform that facilitates communication,which converts text information into sign language gestures through computer technology.Aiming at the problem of low recognition rate of sign language due to the single feature extracted by traditional classification methods or the simplicity of classifier structure.This paper combines the deep convolution neural network architecture as a classifier with multi-feature fusion algorithm,and achieves effective recognition by combining texture features with shape features.In order to avoid over-fitting,the "dropout" method is used to train the deep convolution neural network.Compared with general sign language recognition methods,this method greatly improves the robustness and classification accuracy of the algorithm,and then completes the task of gesture recognition efficiently.To solve the problem that CNN can't deal with continuous time domain and spatial domain related sequence sign language images,A dynamic continuous sign language recognition algorithm was proposed,which combined CNN and RNN.Firstly,according to the extraction method of static sign language recognition,the features were extracted and combined.Then,the one-dimensional vector was used as the input of the Gated Recurrent Unit(GRU)type RNN network,which converted by the continuous two-dimensional feature image,.The DCNN-GRU network identified,designed train methods and configured experimental data,then continuously modified parameters toachieve optimal identification and comparison.Experiments show that the proposed method has higher recognition rate in continuous sign language recognition.