Research On Deep Learning Gesture Recognition Based On Transform Domain Measurement

Human-computer interaction has been a very popular research topic in the computer field,and gesture communication is one of the most commonly used communication methods in people’s daily life,so gesture recognition has become an important research direction in human-computer interaction.At present,gesture recognition has been applied in many industries,and the main technologies include non-visual-based gesture recognition technology via data glove,and visualbased gesture recognition technology via image.In this paper,we study single-pixel detectionbased gesture recognition technology based on transform domain measurement and deep learning,which has the characteristics of less data collection and faster processing.Based on the in-depth discussion of the way of base-pattern generation in the transform domain measurement method,the paper calculates the product of gesture image and orthogonal transform base-pattern as gesture transform domain coefficients,and replaces the pixel information of gesture image with gesture transform domain coefficients,so that fewer transform domain coefficients can be used as feature data set while retaining the gesture image category information to the maximum extent.Then,according to the characteristics of the gesture transform domain coefficients,a suitable deep learning neural network is designed for gesture recognition.In terms of transform domain,considering the difference of different transform domains,four transform domain measures of discrete cosine transform,discrete Hartley transform,Hadamard transform and Fourier transform are selected with the gesture image data to calculate the corresponding gesture transform domain coefficients datasets.Convolutional neural networks and recurrent neural networks are used to train the gesture transform domain coefficients datasets respectively,and the corresponding network models are derived.The test results of the simulation experiments show the feasibility of the proposed method.The experimental platform of singlepixel detection is built,and the gesture recognition work is carried out with the trained neural network models.The experimental results show that the deep learning gesture recognition based on Fourier transform measurement has the highest accuracy,in which the average recognition accuracy is 80% under the condition that the number of coefficients in the transform domain is 35 with the convolutional neural network method,and 62% under the condition that the number of coefficients in the transform domain is 50 with the recurrent neural network method.