Military Sign Language Recognition Based On Depthwise Separable Network

Military sign language is an important method of tactical communication,especially when distance is limited or silence is required.Unfortunately,when soldiers can’t see each other,the use of military sign language is no longer effective,which may delay military operations.In recent years,vision based gesture recognition method has been in the forefront of the field of gesture recognition,but there are still challenges and difficulties in the research of gesture recognition.First of all,the tactical gesture in military sign language is a dynamic sequential action,which is completely represented by the appearance of the hand and the trajectory of the hand movement.How to design the hand shape and trajectory features that can fully describe the characteristics of the tactical gesture is a problem that must be solved in this research topic.Due to the flexibility and details of military sign language,as well as the strong sequential requirements,it brings some challenges to the recognition accuracy.In addition,there is still a lack of specific data set and model for military sign language recognition.In view of the above problems,this paper studies the deep learning technology,and proposes a method to solve the difficulties in military sign language recognition,The main contents and innovations are as follows:(1)Through literature research,this paper investigates the current research status of military sign language recognition,then summarizes the current research status at home and abroad in the field of gesture recognition,as well as the theory and practice of existing methods.(2)A new first person gesture dataset called MSL is made,which contains 16 types of tactical gestures in military sign language,a total of 3840 real samples,more than110000 video frames,and the size is 320×240.The data set has good complexity and authenticity,and can be used for deep neural network training to solve the lack of specific public dataset in current military sign language recognition research.(3)Aiming at the sequential requirement of military sign language recognition and the difficulty of gesture feature extraction,this paper introduces 3D convolution network,improves the basic 3D convolution network C3 D,and constructs the 3D residual network architecture.Through the independent learning ability of 3D convolution for spatiotemporal features,instead of human-designed motion features,the spatiotemporal features of military sign language data are extracted to realize the accurate recognition of military sign language,which verifies the effectiveness of the method.(4)This paper proposes a depth network based on depth separable convolution,named St xception architecture,which extends convolution kernel and pooling kernel to three dimensions.The network can describe the inherent Spatio-temporal relationship of tactical gestures.By deploying deep separable convolution,the number of parameters needed in the network is significantly reduced,and the efficiency of 3D convolution network model is greatly improved.The adaptive average pooling is used to replace the fully connected layer,which further reduces the amount of calculation and the over fitting problem.The results of the experiment show that the proposed model is superior to the existing models on the MSL dataset and other two benchmark datasets.