AI-Signaturer: AI Automatic Signature Generation Based on Pix2pix

Abstract

This paper explores AI automatic signature generation technology based on the Pix2pix conditional generative adversarial network (cGAN), aiming to solve the issues of low efficiency and poor consistency associated with traditional manual signing. Leveraging the learning capability of deep learning models, this study is dedicated to achieving automatic, efficient, and high-quality signature generation, providing technical support for areas such as electronic documents. For this purpose, the paper has designed and trained a Pix2pix model consisting of a generator and a discriminator that can convert input images into signature images with specific font styles. In experiments, 900 signature samples collected manually by students were used for training. To optimize model performance, strategies such as early stopping and model checkpoints were introduced, and hyperparameters like learning rate and iteration counts were adjusted during training. Ultimately, the model achieved an average accuracy of 76% across three different types of input fonts—Kaiti, YaHei, and FangSong—and clearly displayed Chinese character features. However, for complex and dense font styles, there is room for improvement in the model's performance. Future research will focus on more refined data preprocessing and exploring more complex model structures like CycleGAN to enhance performance in multi-font recognition tasks, improving the model's generalization and robustness.