| The style transfer of Chinese characters is a complex and challenging research problem.Different from English,Chinese characters are more complicated in structure,stroke and style.The generation of supervised learning against the network requires a lot of symmetric data,which is very difficult for the collection and collation of many handwritten Chinese character fonts.In addition,the generation of unsupervised learning against Chinese fonts generated by the network will have more stroke errors,and even unrecognizable.Moreover,in previous studies,the style transfer generation confrontation network only focused on the style of the generated Chinese font,while ignoring the resolution of the generated Chinese font.Therefore,this paper proposes SR-Cycle GAN to improve the integrity of the generated fonts and the resolution of the generated font images.Unsupervised generation of adversarial networks uses non-matching data for training,which can save a lot of collection and sorting of Chinese character font images.However,when the generation of unsupervised learning is against the network's style transfer,problematic fonts such as missing strokes,errors,and blurs are generated.Although the normalization operation can speed up the network training speed and the stability of the network training,the batch normalization operation will lose more image detail information.In this regard,reducing or changing the use of the batch normalization layer allows the generation network to learn more style details.This article introduces TV loss to restrict noise and reduce the gap between two pixels.By reducing TV loss,the generation network can generate handwritten Chinese fonts with relatively complete strokes.Because the original Cycle GAN can not generate higher resolution images.Therefore,this paper improves the original Cycle GAN network and increases the number of upsampling layers in the decoding network to enable the generation network to generate images with higher pixels.The original Cycle GAN used transposed convolution to perform upsampling.Transposed convolution did not perform well in super-resolution reconstruction.For this,the transposed convolution was changed to subpixel convolution for upsampling.At this time,the image pixels of the Chinese font generated by the forward mapping are n times larger than the original,and the image pixels of the Chinese font generated by the reverse mapping are also n times larger.According to the principle of cycle consistency of the algorithm,after a cycle,the pixels of the original font will be enlarged to the square of n,so as to achieve the effect of super resolution.This paper uses 240 training sets in HWDB1.1 to train Tensorflow and Chinese handwritten Chinese character recognition models,and 60 test data set test models to obtain Top-1 84.6% and Top-3 93% recognition accuracy.Use Lanting font and HW-1252 as style fonts,italics and bold fonts as conversion fonts,train SR-Cycle GAN to get the generated style fonts and super-resolution reconstruction fonts.Randomly select some samples from the 3,755 generated and reconstructed handwritten Chinese character fonts.Try to ensure that the selected samples must have both simple and complex fonts.The selected samples are input into the trained Chinese handwritten Chinese character recognition algorithm.The average recognition rate of the original Cycle GAN is 33.4%,and the average recognition rate of the SR-Cycle GAN is 92%.The generated and reconstructed Lanting fonts also perform well.The reconstructed Lanting font Vololla has increased by 455,144,262,the brenner has increased by 25,153,766,the SMD has increased by 413,252,and the variance has increased by 482,932,680.The reconstructed handwritten Chinese character image Vollath increased by 67,953,839,brenner increased by 1,178,850,SMD increased by 137,866 and variance increased by 76,186,361.It can be concluded that SR-Cycle GAN has better performance in handwritten Chinese character font style transfer and super-resolution reconstruction. |
PDF Full Text Request