Research On Gesture Recognition Algorithm Based On Improved CNN And SVM

With the progress of modern science and technology,as well as the way of life is more and more intelligent,the demand for natural interaction is becoming higher and higher.Therefore,in view of this phenomenon,it is of great significance to research a natural and comfortable human-computer interaction.Gesture which is natural,intuitive,concise,humanized and flexible.It can fully stimulate the potential of the hand,which do not rely on the mouse,keyboard and other external devices,and avoid limitations.It can be used for human-machine interaction.In recent years,the convolution neural network model has been widely used in various fields of classification research.Especially in the field of image recognition,it can improve the accuracy of image recognition,and is superior to many traditional methods in feature extraction.However,the training of convolutional neural network requires large amounts of dataset and large computation.Training on small datasets is prone to overfitting or low recognition accuracy.But SVM can achieve high accuracy recognition on the basis of effective feature extraction.In this paper,a method based on convolution neural network(CNN)and support vector machine(SVM)is proposed for small dataset and applied to gesture recognition.The main research work in this article is as follows:(1)Convolution neural network can extract multi-scale features,By optimizing the architecture layer number,and add Dropout and Regularization constraint to the full connection layer which improve the generalization ability of the full link feature.(2)Using SVM classifier instead of Softmax classifier in convolutional neural network,the connection layer after feature extraction using SVM classifier for training,get a balanced and sparse recognition model,So CNN+SVM classifier can achieve good generalization performance in small datasets.(3)The model is applied to gesture recognition,and the model is verified on the Sebastien Marcel and Jochen Triesch gesture dataset.The recognition accuracy is 98.7% on the Sebastien Marcel gesture dataset,and the recognition accuracy is 98.6% on Jochen Triesch gesture dataset.At the same time,the paper compares with the four methods of Improved CNN?CNN+KNN?CNN+ AdaBoost and CNN+Random Forest.The results show that the accuracy of CNN+SVM classifier on two gesture datasets is higher than that of the others.