Artistic Text Effect Transfer with Conditonal Generative Adversarial Network

With the wider application of artistic text in real scenes, the artistic text effect transfer has become an increasingly important topic in the field of computer vision. Existing methods either maintain poor structural information or poor effect information, both of which can seriously deteriorate the final result. We propose a conditional generative adversarial network to solve this problem, by alternately training the generator and the discriminator. The structure of the generator is Unet, which is an encoder-decoder with skip connections between the corresponding mirrored layers. We inject the conditional information of the reference effect image into the generator using a simple but effective concatenate strategy. The discriminator, which is conditioned on the generator's input to determine how true its output is. Ideally, the discriminator acts as a loss function and guides the optimization of the generator. Qualitative and quantitative results are performed on TE141k dataset, which demonstrate the proposed network outperforms the existing SOTA models and has good visual results.