Research On Hand Gesture Recognition Based On Convolutional Neural Network

Because gesture is simple,direct and easy to understand,people use different methods to study gesture recognition and its application in different periods,making it an important human-computer interaction mode.For example,in the aspect of flight attitude control of UAV,the ground uses camera to collect gesture images,and then controls UAV by recognizing and judging the meaning of gesture.It has the advantages of flexible and convenient control,so it has good application value.Traditional image recognition method relies on manual design of complex feature extraction algorithm for image feature extraction.In recent years,great progress has been made in building convolutional neural network model and using deep learning algorithm to realize image recognition,showing great advantages.Compared with the traditional image recognition method,the convolution neural network can directly take the whole image as input,extract the features through the convolution layer by layer inside the convolution neural network,predict the specific categories of gestures by classifier,and optimize the parameters of the convolution neural network by using back propagation algorithm according to the functional relationship between the predicted value and the real value.Thus,the convolution neural network can be used to optimize the parameters of the convolution neural network.It can extract image features from different aspects,which makes convolutional neural network have excellent image recognition performance.It avoids the singularity of feature extraction by artificial design feature extraction algorithm,and greatly improves the recognition rate and robustness of image.The main work of this paper is to construct an improved convolutional neural network model and improve the recognition rate and robustness of gesture images by using deep learning algorithm.The main research work completed is as follows:(1)In order to solve the problem of low recognition rate and easy to be disturbed by the surrounding environment in complex background,a compound convolution neural network algorithm is proposed to realize gesture image recognition.When recognizing gestures,because there will be different background interference in different scenes,skin color interference is a more influential interference,because the skin color of the hand is the same as that of the face.In order to improve the performance of the proposed gesture recognition algorithm,this paper collects a large number of data sets of gesture image production which are mixed with face and skin color in complex environment.The coordinates of gestures in the image are marked by the annotation software,and the gestures and background are distinguished.(2)Construct the convolution neural network model,train the convolution neural network with the set of gesture image data,and then use the trained convolution neural network for gesture detection.In order to improve the accuracy of gesture detection of convolutional neural network,the method of segmentation of the detected gesture based on gesture area algorithm is studied,which can ensure that the detected gesture image is a complete gesture and reduce the interference of the surrounding environment on gesture recognition.(3)In order to solve the problem of low recognition accuracy when using convolution neural network for gesture recognition,this paper firstly preprocesses the gesture image detected by single convolution neural network.Through filtering,background elimination and binarization,the gesture image is highlighted and the background image is weakened,so that the influence of background on gesture recognition is reduced.Secondly,based on the classical convolution neural network model Alexnet,a convolution neural network structure with two improved multi-scale convolution kernels is proposed.The first one is to improve the first convolution layer of convolution neural network.This improved convolution neural network model adopts two convolution kernels to extract gesture features from gesture images in the first convolution layer.The second one is to continue to use multi-convolution kernels to extract gesture features from deep convolution layer on the basis of the first improvement,using multi-scale convolution.Kernel and two channels are used to fuse gesture features,and then the improved model is validated by simulation experiments using the segmented original gesture image data set and binary gesture image data set.The simulation results show that the recognition rate of the improved convolutional neural network is obviously improved,and the correct recognition rate of gesture image is 95.64%.