Seg-CycleGAN: An Improved CycleGAN For Abstract Painting Generation

In modern art, the creation of abstract painting has become a prominent artistic expression with significant aesthetic merit. With the development of artificial intelligence, the abstract paintings generated by AI has gradually been recognized by the public. This paper proposes an abstract painting creation method based on an improved CycleGAN network (Seg-CycleGAN). Our method starts by obtaining a collection of color blocks through clustering and color block reorganization based on the original input image which is then fed into the CycleGAN network for style transfer. In addition, we add a clustering layer on one of the generator of CycleGAN, which can effectively accelerate the convergence speed of the generator and improve the training efficiency and quality of results. This work provides a novel method for AI generation of abstract paintings and demonstrates its effectiveness in experiments.