Design Of Gesture Recognition Algorithms Based On Deep Learning

Gesture recognition technology is a kind of biometrics technology to recognize people's hand posture.The traditional gesture recognition algorithm based on machine vision has weak anti-interference ability and low recognition accuracy,so it has little value in practical application.Compared with the traditional gesture recognition algorithm based on machine vision,the gesture recognition algorithm based on convolutional neural network has better anti-interference ability and recognition accuracy,but there are still some problems.Aiming at the problems of large network model and low recognition speed and accuracy in gesture recognition algorithm,this paper analyzes gesture recognition algorithm based on convolutional neural network,optimizes SSD network in gesture recognition algorithm,and improves the practical effect of gesture recognition algorithm.This paper mainly completed the following work:(1)The front part of the SSD network in the gesture recognition algorithm of this paper is optimized,which solves the problem that the SSD network model is too large and the recognition speed is slow in the gesture recognition algorithm.This paper optimizes the network structure of the traditional MobileNet.First,it intercepts the first 6 layers of traditional MobileNet convolution and adds 4 layers of new deep separable convolution to form the optimized MobileNet,and then replaces the gesture recognition algorithm with the optimized MobileNet.The original front part of the SSD network improves the traditional SSD network.In this paper,the improved SSD network is compared with the traditional SSD network in the gesture recognition algorithm.The improved SSD network reduces the network model size of the gesture recognition algorithm by 55.3%and the recognition speed by 36.1%.(2)The multi-dimensional feature structure of the SSD network in the gesture recognition algorithm of this paper is optimized to solve the problem that the SSD network can not fully utilize the top-level features.In this paper,the top-level features with strong semantic information and the features of the next layer are merged to form the first fusion feature layer.Then the new feature layer is regarded as the top-level feature layer and the next feature layer.Fusion to form a second fusion feature layer;then repeating the above steps to obtain a third fusion feature layer;finally,using all the fusion feature layers and the original feature layer of the multi-dimensional feature structure of the SSD network to form a new multi-size feature structure and The structure is applied to the improved SSD network in(1),and the SSD network is optimized.This paper compares the improved SSD network with the traditional SSD network in the gesture recognition algorithm.The improved SSD network makes gestures.The recognition algorithm's mAP increased by 6.3%.(3)Application of depth separable convolution and hole convolution replaces the standard convolution of the auxiliary convolutional layer of the SSD network in the gesture recognition algorithm of this paper,and optimizes the improved SSD network in(1).Among them,the use of depth separable convolution can reduce the amount of calculation and the number of parameters of the network,and the use of cavity convolution can increase the receptive field of the feature.This paper compares the application of the optimized SSD network with the traditional SSD network in the gesture recognition algorithm.The optimized SSD network reduces the network model size of the gesture recognition algorithm by 62.6%and the recognition speed by 59.0%.