Research On Handwritten Font Generation Algorithm Based On Generative Adversarial Network

In recent years,with the revival of deep learning,computer vision has become a hot research topic of artificial intelligence.It focuses on how technology can enable computers to automatically acquire high-level understanding from a range of number images or video in an attempt to create artificial intelligence systems similar to human visual systems.Nowadays,its application scenarios have been extended to various aspects such as industrial,agricultural,medical,and transportation,furthermore automatic quality inspection,automatic robot control,face recognition,fingerprint recognition,and automatic driving are included.However,most of these applications use a supervised learning-based model architecture,which cannot meet the increasingly large and complex application environment in the future.Therefore,attempts to solve the problem of computer vision by unsupervised learning have attracted much attention.As the most specific and typical application in the field of computer vision,virtual reality technology uses computer to generate realistic visual,auditory,tactile and other sensations,so it enables users to interactively experience real dynamic scenes in a virtual environment.In order to get the best visual experience,the quality and variety of image generation in the scene is particularly important.Aimed to this purpose,this thesis focuses on the problem of image generation modeling,and deeply analyzes the related structure design of the basic Generative Adversarial Network which produces diversified handwritten number body images.To solve the problem that the number outline of the generated handwritten picture is not sharp enough,the Convolutional Neural Network is applied to improve the structure of the generation network and the discriminant network in Generative Adversarial Network respectively.The Convolutional Neural Network can share the weight,which greatly reduces the parameters of the model.In return,the quality of the generated handwritten number body images gets much better.In this thesis,the handwritten number pictures generated by the variational auto-encoder model,the basic Generative Adversarial Network model and the improved Generative Adversarial Network model are further compared and analyzed.The evaluation criteria of the generated pictures are discussed in order to enhance the interpretability of the picture generation model.The improved model in this thesis has good generalization ability and outstanding adaptability to fresh samples,which further highlights the advantages of Generative Adversarial Network in unsupervised learning.In this thesis,Python 3.6.5 language and TensorFlow 1.8.0 architecture are utilized for model training,also MATLAB2017 a tools is used to assist the implementation of the algorithm.