Research On Gesture Recognition Algorithm Based On Convolutional Neural Network

Today,smart devices such as computers have entered millions of households,and human-computer interaction with smart devices has gradually become an essential part of people's daily lives.Human-computer interaction mainly controls and operates computers or smart devices through information conversion between humans and computers to achieve functions and purposes.Human-computer interaction has received widespread attention in terms of sign language recognition,virtual reality,device control,and entertainment.Gesture interaction is widely concerned by researchers because of its convenience and the rapid and convenient acquisition of gesture images by smart phones and other devices.Research on gesture interaction based on computer vision has become a hot topic in the field of human-computer interaction.However,gesture images contain complex and diverse background noises,such as lighting and background overlap occlusion.Gesture recognition based on computer vision still faces huge challenges.The emergence of machine learning and deep learning provides new ideas for gesture recognition based on computer vision.In this thesis,the problem of gesture recognition based on computer vision is studied in depth,and a gesture recognition method based on convolutional neural network is proposed.Accurate and fast gesture recognition under complex background is realized.The research in this article mainly includes the following parts:(1)In order to reduce the influence of complex background,lighting and other noises on recognition performance in gesture images,a DSSD-based gesture image detection algorithm is proposed.First,based on the VGG network and adding convolutional layers for feature extraction,build an SSD-based target detection model,and then use the data set to train and test the monitoring model.In order to improve the accuracy of small target detection,a combination of shallow and deep feature information is proposed,and a DSSD-based gesture detection algorithm is proposed.Experimental results show that the model can effectively detect gesture areas.(2)In order to segment gestures from the background image to improve the accuracy of gesture recognition,a gesture image segmentation algorithm based on Generative Adversarial Networks(GAN)is proposed.First,the AlexNet network is fully convolved,and then a hop-level structure is combined with multi-layer features to perform deconvolution to establish a gesture segmentation algorithm based on a Full Convolutional Neural(FCN)network.Then,aiming at the problem of loss of details after segmentation of the gesture image caused by the upsampling of FCN,and referring to the idea of GAN,a gesture image segmentation algorithm based on generative adversarial network was proposed.Finally,the performance of the gesture image segmentation algorithm is evaluated on a public data set.The experimental results show that the gesture image segmentation algorithm based on GAN can remove background and noise areas and effectively extract gesture areas.(3)In order to achieve effective gesture recognition,a gesture classification model based on GoogLeNet network is proposed.First,a segmented gesture image is used to build a training set,and the recognition model is trained.In the test phase,a segmented image is input into the recognition model,and its output is the category of the image.In the experiment,the original gesture image,gesture detection image,and gesture segmentation image were used to train the GoogLeNet network to obtain three gesture recognition models,and then the recognition results of these three models were analyzed and compared.The experimental results show that the recognition model based on GoogLeNet has good stability,can extract the robust features of images under complex background,and improve the accuracy of gesture recognition.In addition,gesture detection and gesture segmentation preprocessing algorithms can improve the accuracy of gesture recognition.