Research On Gesture Recognition Method Based On Improved Convolutional Neural Network

With the vigorous development of emerging information technologies such as Artificial Intelligence,the Internet of Things,5G,and Cloud Computing,Human Computer Interaction has gradually become intelligent and diversified.Gestures are widely used in real life such as robot control,virtual reality,intelligent driving,and handheld pan-tilt due to their strong information expression ability and transmission function,simple and easy to understand,non-contact,and other characteristics.However,gesture recognition also has some problems such as the high similarity between gesture categories and self-occlusion between fingers,which leads to low accuracy of gesture recognition.At the same time,the gesture recognition algorithm based on the convolutional neural network has a large number of parameters and requires high computer hardware,and it is difficult to apply to mobile and embedded devices with limited resources.Therefore,the research on gesture recognition methods has important significance and practical application value.Based on the theoretical basis of the convolutional neural network,this paper adopts a lightweight convolutional neural network model.By rationally designing the structure and parameters of the network,while reducing the complexity of the model,it improves the accuracy of gesture recognition.The main research work and innovations completed in the thesis are as follows:1.A dual-channel feature fusion gesture recognition network with a weighted loss function is designed(WDN)Aiming at the problems of low accuracy of gesture recognition and a large number of model parameters,this paper designs a dual-channel feature fusion gesture recognition network with a weighted loss function.First of all,the network is a parallel dual-channel structure.The local features of gesture images learned by two parallel sub-networks are merged to obtain richer gesture feature information.Next,to balance the weights of the local features learned by the two channels,a weighted loss function is used in the training process,which helps to obtain better gesture recognition accuracy.Then,to reduce the number of model parameters,the last layer of the network uses a convolutional layer instead of a fully connected layer for classification.The experimental results show that the dual-channel feature fusion gesture recognition network with weighted loss function designed in this paper effectively improves the accuracy of gesture recognition and has fewer model parameters.2.A dual-channel feature fusion gesture recognition network with an attention mechanism and residual connection mode is proposed(ARDN)To further improve the accuracy of gesture recognition and reduce the number of parameters of the model,the residual connection method,attention mechanism,and deep separable convolution technology are introduced into the dual-channel network.First of all,the proposed residual connection method uses residuals and image decomposition knowledge to learn gesture feature information more accurately.Then,the attention mechanism is introduced to automatically obtain the importance of each channel in the gesture image feature map,and according to the importance degree,the useful features are enhanced and the useless features are suppressed,to realize the adaptive weighting of the feature channels.Finally,using depth separable convolution instead of ordinary convolution can greatly reduce the number of model parameters.The experimental results show that the dual-channel feature fusion gesture recognition network of the attention mechanism and residual connection method proposed in this paper further improves the accuracy of gesture recognition and reduces the number of model parameters.