By adding residual connections, each layer in the generator depends on the previous layer's output, directly transmitting information through shortcut paths, and maintaining the stability of the network in deep training. Combined with style transfer technology, the artistic style is integrated with the font structure further to enhance the personalised characteristics of the generated font. Experiments show that the method in this paper significantly improves the quality, detail expression, and customised style of the generated fonts. The structural similarity index (SSIM) of the generated font image and the target image pixel block is between 0.82 and 0.93, and the recognition accuracy of the generated font in the optical character recognition (OCR) model is not less than 0.7. The details of the generated font are very close to the target image, effectively making up for traditional generative adversarial networks' shortcomings in art font design.
Yin et al. (Thu,) studied this question.