Research On Gesture Recognition Based On Structural Features And Deep Learning

In recent years,with the rapid progress of artificial intelligence technology,the development of flexible and convenient intelligent human-computer interactive mode is increasingly urgent.Gesture interaction is intuitive,natural,rich,consistent with human communication habits characteristics,which has potential applications in the field of industrial control,intelligent home furnishing and sign language recognition.Hand gesture recognition as the key technology is caused by domestic and foreign scholars paid more and more attention.A gesture recognition method with high recognition rate and good stability is proposed by using gesture structure features and deep learning method.The method includes image preprocessing,feature extraction,classification and recognition,and the specific contents are as follows:The gesture preprocessing method based on skin color,shape and contour information was studied.Firstly,skin color segmentation image was obtained by combining skin color information and Bayesian model in YCrCb space;Secondly,denoising and shape segmentation were carried out by using the information of the area and contour length of connected domain and gesture palm region in skin color segmentation image,and the gesture target region was obtained.finally,the direction of gesture target region was corrected by using Hough transform to detect the straight line of contour.The methods of gesture feature extraction including contour structure,finger structure and texture structure were studied.The main contents are as follows:(1)Extracted the gesture contour information of geometric and time domain and frequency domain of the unfolding curve,contour structure characteristic was constructed;(2)Near-convex decomposition algorithm was designed to obtain direction,position and shape information of finger in gesture to construct finger structure feature;(3)Gray-level Co-occurrence Matrix(GLCM)features and Gabor features were extracted from spatial and frequency domain information to construct texture structure features;(4)Gesture structure features were formed in series after normalization of contour,finger and texture structure features,so as to improve the utilization of gesture category information in images.The classification and recognition method of gesture structure features for deep learning was studied.Firstly,a deep stack encoder network was constructed by using sparse automatic encoder and Softmax classifier in deep learning.The network adopts four layers of structure: input layer,two hidden layers of sparse automatic encoder and output layer of Softmax classifier.Then,gesture structure features were used as input of deep stack encoder network to carry out input characteristics for deep learning and classification recognition.The experimental results show that the gesture recognition method achieves the segmentation of gesture target area and the extraction of gesture structural features,and can use the extracted features for gesture recognition.The recognition rate reaches 97.7%,which is higher than other existing methods.It has better invariance of illumination,rotation and zoom,and achieves the expected goal of the study.