GL-GAN: Adaptive global and local bilevel optimization for generative adversarial network

Highlights：

• With a simple network structure, GL-GAN is allowed to effectively avoid the nature of imbalance by global and local bilevel optimization, which is accomplished by first locating low-quality areas and then optimizing them.

• By using feature matrix cues from discriminator output, we propose the adaptive local and global optimization method(Ada-OP) for specific implementation and find that it boosts the convergence speed.

• Compared with the current GAN methods, the model has shown impressive performance on CelebA, Oxford Flowers, CelebA-HQ and LSUN datasets.

摘要

•The model achieves the generation of high-resolution images in a complementary and promoting way, where global optimization is to optimize the whole images and local is only to optimize the low-quality areas.•With a simple network structure, GL-GAN is allowed to effectively avoid the nature of imbalance by global and local bilevel optimization, which is accomplished by first locating low-quality areas and then optimizing them.•By using feature matrix cues from discriminator output, we propose the adaptive local and global optimization method(Ada-OP) for specific implementation and find that it boosts the convergence speed.•Compared with the current GAN methods, the model has shown impressive performance on CelebA, Oxford Flowers, CelebA-HQ and LSUN datasets.

论文评审过程：Received 18 November 2020, Revised 9 September 2021, Accepted 15 October 2021, Available online 20 October 2021, Version of Record 7 November 2021.