Research On Gesture Detection And Recognition Based On Deep Learning

Gestures are a multi-faceted communication method that plays an important role in non-verbal communication and human-computer interaction.They provide an attractive solution for human-computer interaction(HCI).Among them,the first interactive method was through external hardware such as digital gloves,biological EMG,kinect depth equipment,etc.,and gradually developed into a method based on computer vision algorithms.The latest developments in deep learning have greatly improved the performance of image recognition.Divorced from hardware devices,it is in line with people's convenient and fast usage habits.However,gesture recognition still has limitations in unfavorable scenes with changes in gestures,changes in lighting,or complex backgrounds.The rapid development of related deep learning algorithms has caused problems in gesture recognition.Has been improved.Therefore,this article combines the relevant knowledge of image processing and deep learning to gesture recognition,realizes the recognition of one-handed and two-handed gestures,improves complex environments,including gesture recognition under different lighting and complex backgrounds,and proposes improvements in image processing Method,improve the efficiency of gesture recognition in all aspects.Aiming at the problem of the scarcity of two-hand gesture recognition algorithms for gesture recognition,the segmentation and recognition algorithm of two-hand gestures is studied,and a method for segmentation and positioning of two-hand gestures based on image processing is proposed.For the lack of two-hand gesture data sets,selfmade two-hand gesture data sets with complex and simple backgrounds.Research the comparison of the recognition efficiency of LBP,PCA and HOG three traditional gesture recognition methods and deep learning gesture recognition methods,and improve the problems of low accuracy,poor convergence,and low robustness of deep learning gesture recognition methods.In convolutional neural Based on the network,an adaptive enhancement module is introduced,which performs adaptive residual enhancement to improve performance based on the results of each network training,and proposes an adaptive enhanced convolutional neural network(AECNN)gesture recognition model,and combines it with the YCBCR skin color model.Gesture extraction improves the efficiency of gesture recognition.The two-handed gesture segmentation algorithm is applied to the self-made twohanded gesture data set,and combined with the five-layer network hand count classifier to realize the gesture grouping prediction.The experiment shows that the correct prediction result is 98.82%.AECNN is applied to the United States(ASL)9 types of single-handed gesture data sets,comparing traditional gesture recognition methods,local binary pattern(LBP),directional gradient histogram(HOG),HOG combined with principal component analysis(HOG+PCA)combined with SVM Realize gesture recognition.Compared with the traditional convolutional neural network CNN in deep learning and the gesture recognition of CNN+Dropout layer,the experimental results show that the average recognition rate of AECNN gestures is as high as 97.87%.Both the recognition rate and convergence are much higher than Traditional gesture recognition methods and traditional CNN.Comparing the recognition performance of each specific gesture,the experiment shows that most gestures can be accurately recognized.The gesture recognition experiment is carried out by adding an appropriate level of noise to the test data and the experiment is carried out by extracting test pictures with different backgrounds and lighting.The experimental results show that the two experiments are average The recognition rates are about 96% and 94%,respectively,and AECNN has good robustness to adverse environmental impacts.