Handwriting Identification And Generation Technology Based On Generative Adversarial Networks

In the field of image processing and pattern recognition,handwriting recognition refers to the automatic recognition of handwriting.When the computer extracts the features of handwriting,it uses classification method to identify handwriting.However,the traditional image recognition methods have some problems,such as lack of standard basis for feature selection and low accuracy of identification.Handwriting generation is a function of machine intelligence,which uses computer to generate handwriting automatically.However,the traditional methods can only generate standard fonts.GAN(Generative Adversarial Nets)is a kind of deep learning model,which generates better output through game learning between the discriminative model and generative model.In this thesis,the problem of handwriting identification and generation is studied by using Generative Adversarial Nets.SIGAN(Signature Identification GAN)network and improved SIGAN network is designed to realize the identification and generation of Chinese hard-pen signatures and soft-pen handwriting.The main work is as follows:1.Build a handwriting images data set,which includes 7060 hard-pen signatures and soft-pen handwriting.There are 2560 hard-pen signatures,including 1280 real hard-pen signatures and 1280 imitative hard-pen signatures.There are 4500 soft-pen handwriting,consisting of 15 major influential mainstream calligraphic styles,including five traditional calligraphic fonts:running script,cursive script,official script,seal script and regular script,each of which has 300 calligraphic styles.2.By using dual learning for reference,a special SIGAN network is designed to realize the task of handwriting identification and generation.The basic GAN is further extended to two coupled GANs,including two generators and two discriminators.In the training process,two GANs play games with each other.After the convergence condition is reached,the training is completed.At this time,the model has the generating function.The loss value of the discriminator is used as the discrimination threshold to achieve the purpose of identifying handwriting.The experimental results show that the average accuracy of hard-pen signatures and soft-pen handwriting identification are 91.2%and 73.2%respectively.The average GAN-test of hard-pen signatures and soft-pen handwriting images are 92.2%and 74.7%respectively.3.By using attention mechanism for reference,an improved SIGAN network is designed to achieve handwriting identification and generation tasks.The attention layer is added between the sixth and seventh convolution layers of SIGAN generator,and the traditional convolution feature map is replaced by the attention feature map to further improve the performance of the model.The experimental results show that the average accuracy of hard-pen signatures and soft-pen handwriting identification are 92.5%and 88.6%respectively,which is 4.9%and 17.3%higher than that of traditional image classification methods.It is also much higher than that of subjective eye tests(72.3%and 79.3%).The average GAN-test of hard-pen signatures and soft-pen handwriting images are 93.6%and 82.7%respectively.