Research On Gesture Recognition Based On Deep Learning

Most smart home systems and AI robots are mainly voice control or APP control,which makes some special groups unable to enjoy the convenience of AI to life.Under this research background,this paper innovates the structure of convolutional neural network algorithm to make it more suitable for artificial intelligence control system.In order to enable these special groups to enjoy the convenience of artificial intelligence to life.The specific work is as follows:(1)By comparing and analyzing various gesture recognition algorithms,the basic framework of gesture recognition based on convolutional neural network algorithm is finally determined.(2)For the special case of smart home,the collected pictures include complex gesture structure,illumination,background,environment and other factors.If not pretreated,the collected pictures will be directly used for training and recognition,which will affect the accuracy of recognition.Therefore,this paper presents a gesture recognition scheme which combines skin color model and convolution neural network.(3)For gesture images with complex background,the following two steps are used to complete the image preprocessing.In the first step,the appropriate skin color model is selected to determine the location of the gesture region in the original image,and the region is separated from the background.The second step is to extract andreconstruct the separated gesture regions using morphological operations,filtering algorithms and connected domain marking.(4)After completing the original image preprocessing,the working principle of convolution neural network is analyzed in detail,the main algorithms in the training process are introduced in detail,and the formulas are deduced.On the basis of the above research and analysis,the network structure,the number and size of convolution kernels,the method and size of pooling,the activation function and the classification method of the convolution neural network are determined through comparative experiments.At the same time,this paper designs two continuous convolution neural networks to recognize two consecutive times in order to prevent misoperation in actual use.