Chinese Font Generation Algorithm with Multi-receptive Field Features
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
维普
万方数据
针对目前字体生成算法对字体细节特征关注不足,导致生成笔画细节粗糙和风格特征模糊等问题,提出一种融合多感受野特征的汉字字体生成算法.针对笔画细节缺失引入多感受野特征金字塔模块(Multi-Receptive Field Feature Pyramid,MRFP)捕获不同尺度的特征,重建笔画细节,在生成过程中既能处理整体字形又能保留细节特征,提高生成一致性;针对冗余特征影响生成结果质量,引用通道和空间重建卷积(Spatial and Channel Reconstruction Convolution,SCConv)代替普通卷积,采用分离重构和分离变换融合的方法抑制空间和通道的冗余,提高对整体字形特征的关注度,加强模型泛化能力.将多感受野特征金字塔模块和SCConv融合于字体生成模型.实验采用10类中文字库训练,两种测试集验证,将算法与CycleGAN,Zi2zi,DG-Font进行比较.结果表明:该方法的生成结果在整体风格和笔画细节上优于其他字体生成算法,在PSNR,SSIM,LPIPS评价指标上均有改善.
Aiming at the problems of rough stroke details and fuzzy style features caused by insufficient attention to detail features in current font generation algorithms,a Chinese character font generation algorithm based on multi-receptive field features was proposed. In view of the lack of stroke details,a Multi-Receptive Field Feature Pyramid (MRFP) module was introduced to capture features of different scales and reconstruct stroke details. In view of the impact of redundant features on the quality of the generated results,Spatial and Channel Reconstruction Convolution (SCConv) was used instead of ordinary convolution,and the fused methods of separation reconstruction and separation transformation were adopted to suppress the redundancy of space and channels. Therefore,the attention to the overall font features was improved and the generalization ability of the model was strengthened. In the experiment,10 types of Chinese character libraries were used for training,the experimental results showed that the generation result of this method was superior to other font generation algorithms in terms of overall style and stroke details. The evaluation indexes of PSNR,SSIM and LPIPS were all improved.
font generationfeature pyramidredundancy featuresGAN
维普
万方数据
