Offline Handwritten Chinese Character Recognition Based On Generative Adversarial Networks

Accurately identifying and extracting offline handwritten Chinese characters from the image can greatly reduce the manpower and material resources required for the Chinese manuscript collation.Due to the complexity and variety of font glyphs and the wide variety of categories,the recognition of offline handwritten Chinese characters has always been a problem of pattern recognition.In recent years,the fast-developing deep learning network has made rapid development in the fields of image recognition,image generation,and target detection.Using deep convolutional neural networks to solve offline handwritten Chinese character recognition is an important way.However,the application of deep learning requires a large number of training samples.Although the current multi-class sample collection of offline handwritten Chinese characters is relatively large,the detection and identification sample set is still very insufficient.There are too many Chinese characters,different forms,and different frequency of characters.Many words are not common.Even if they are seen,their glyphs are often single,and the real samples with uniform frequency distribution are collected.Therefore,it is difficult to obtain enough quantity samples by manually marking existing documents.Aiming at this situation,this paper will apply Conditional Generative Adversarial Networks(GAN)on the generation of offline handwritten Chinese characters,obtain a large number of offline handwritten Chinese character samples,and then synthesize a sufficient amount of Chinese character detection and recognition samples,and design the detection and recognition algorithm of offline handwritten Chinese characters.This paper proposes a set of offline handwritten Chinese character sample generation,Chinese character detection,and Chinese character recognition methods.Firstly,aiming at the feature extraction and recognition of images,this paper proposes an improved Inception convolutional neural network structure and builds an excellent Joint-Net convolutional neural network classification model.Then,the Joint-Net convolutional network is applied as a discriminant network in the WGAN-GP,and an offline handwritten word sample is generated.Using the generated offline handwritten Chinese characters and the existing offline handwritten Chinese character dataset,a complex background offline handwritten detection detection sample is synthesized.Finally,with the joint-Net as the backbone of the network,a novel end-to-end offline handwritten Chinese character detection and recognition model was designed.In the experimental part,the Joint-Net model was experimentally verified on the offline handwritten Chinese character dataset.Using the stochastic gradient descent optimization algorithm,the CNN model achieved an average accuracy of 96.95% on the test set.The Chinese character detection and Chinese character recognition algorithms are excellent.The experimental results show that off-line handwritten Chinese character recognition based on GAN is a good solution.