Channel and constraint compensation for generative adversarial networks

Abstract

In this paper, we propose channel and constraint compensation mechanism applied in Generative Adversarial Networks (GANs) to help distribution fitting and improve the visual quality of generated samples. The proposed channel compensation focuses on specific featurerelated regions by weighting the channel of conv-layer feature maps, so specific feature modes of data distribution can be compensated and irrelevant features can also be decayed. By combining the Jensen-Shannon (JS) divergence and Wasserstein distance (WD) into a well-designed objective function, the constraint compensation can impose more useful constraints upon the generator to diversify the estimated density in capturing multi-modes. Extensive experiments are conducted on synthetic 2D data and real-world datasets (CIFAR-10, STL-10, CelaBa). The qualitative and quantitative comparisons against baselines demonstrate the effectiveness and superiority of our method.