Research On Generation And Recognition Methods And Applications Of Handwritten Characters In The Air

Nowadays,the aerial handwriting recognition technology has made some achievements,but it is still in the initial stage of development and still needs further research.In order to solve the existing problems of aerial handwriting recognition,realize a fast,stable,safe and easy to expand aerial handwriting recognition method,and meet people’s needs for more natural,more flexible and more intelligent humancomputer interaction,we have conducted in-depth research on aerial handwriting recognition technology,and proposed a new aerial handwriting character recognition system.This thesis mainly makes the following contributions to the development of the recognition technology of freehand writing:(1)In order to solve the problems of low speed,lack of stability,and lack of security and expansion in application of existing aerial handwriting recognition methods.We propose a new aerial handwritten character recognition system.The system combines a lightweight target detection model,a single target tracking model and a character recognition model to complete the task of aerial handwritten characters and character recognition,and realizes a fast,stable,safe and easy to expand humancomputer interaction mode.(2)In order to quickly and stably draw aerial handwritten characters,we studied the single target tracking algorithm in aerial handwritten character recognition system.By tracking hand cards with a single target tracking model,the trajectory of handwritten characters in the air is drawn.Two improvement strategies are proposed for single target tracking algorithm.First,in order to improve the tracking accuracy,a feature extraction network is designed based on receptive field theory to complete the extraction of deep features.Then,in order to optimize the network and improve the reasoning speed of the network,a separable deep convolution is used in the backbone network.(3)In order to recognize handwritten characters in the air and expand the types of character recognition,a recognition method of handwritten characters in the air is designed.In this method,a preprocessing method for aerial handwritten characters and two character recognition models are proposed.Firstly,the slanted characters written in the air are corrected by using the preprocessing of handwritten characters in the air,and the track coordinates of handwritten characters in the air are normalized to solve the problem of different input image scales of different character recognition models.Then,a parallel convolution neural network is proposed for the recognition of handwritten single character in the air.The network improves the recognition rate of a single character by fusing the features of different receptive fields.Finally,a fusion recurrent neural network is proposed for continuous characters written in the air.In the fused recurrent neural network,a feature extraction network composed of fusion modules is proposed,and the number of fully connected neurons is controlled to reduce the parameters and computation of the prediction network.(4)In order to apply the aerial handwritten character recognition system to specific practice,the system is deployed to the embedded platform to verify the portability of the system.On the embedded platform,the whole system is tested.From the test results,it can be found that the system meets the requirements of fast and stable.In addition,in order to verify the expandability of the system,two simple applications are designed on the basis of the system,namely air simple calculator and air simple mail.The aerial simple calculator can complete the aerial handwritten single digit and the number written by operation,and give the operation results.Air simple mail can complete the writing and sending of air simple mail.