Research On Gesture Recognition Based On Deep Learning

At present,the role of human-computer interaction in people's daily life has penetrated into every corner of human production and life,and has also triggered extensive research by scholars at home and abroad.The rapid development of computer software and hardware technology and the continuous upgrading of new algorithms in the fields of gesture recognition have continuously improved the experience of human-computer interaction.Gesture recognition has a very broad application value,and there are a large number of requirements and realizations in the fields of virtual reality,sign language recognition,polygraph,remote communication,driving safety,and medical surgery.In view of the fact that most traditional gesture recognition technologies are based on hardware devices such as data gloves,visionbased gesture recognition technologies that rely less on devices and have low image quality requirements have attracted more and more attention,and deep learning-based gesture recognition technologies have received a lot of attention.Researched and has been applied.This article reads and analyzes relevant domestic and foreign documents,and elaborates the current research status at home and abroad.At the same time,the article also studied the related methods of gesture recognition,analyzed the advantages and disadvantages of mainstream methods,and proposed some research significance of the improvement of this method.Based on the advantages of deep learning in the field of vision,this thesis mainly studies classical convolutional neural networks such as lenet,which constitutes the key work of this thesis.Based on the advantages and disadvantages of the traditional Le Net network learning algorithm,this thesis makes relevant improvements,optimizes a general algorithm,uses the convolution of multiple low channels to replace the convolution of a single high channel,and improves the model representation ability and running speed.The algorithm uses two groups of equally sized convolution cores for grouping convolution,convolutes on multiple groups at the same time,and aggregates different groups at the same time to get good feature information.Through 1500 rounds of training data,the expected goal is achieved,the accuracy rate is greatly increased,and the loss function is greatly reduced.Through experiments,the test accuracy rate of the algorithm proposed in this thesis reaches 96.4%,and it shows better classification and recognition ability when running on a self-built database,which improves the recognition rate.