Adaptive adversarial training method for improving multi-scale GAN based on generalization bound theory
This work addresses the problem of improving multi-scale GAN performance for image processing tasks, offering a novel training approach that enhances image quality with significant metrics gains, though it is incremental in applying theoretical bounds to a specific model type.
The paper tackled the convergence and capability limitations of multi-scale GANs by introducing PAC-Bayes generalization bound theory to analyze training, leading to an adaptive adversarial training method that improved image manipulation quality. For image super-resolution, it achieved a 100% reduction in NIQE and 60% reduction in RMSE, outperforming models trained on large datasets.
In recent years, multi-scale generative adversarial networks (GANs) have been proposed to build generalized image processing models based on single sample. Constraining on the sample size, multi-scale GANs have much difficulty converging to the global optimum, which ultimately leads to limitations in their capabilities. In this paper, we pioneered the introduction of PAC-Bayes generalized bound theory into the training analysis of specific models under different adversarial training methods, which can obtain a non-vacuous upper bound on the generalization error for the specified multi-scale GAN structure. Based on the drastic changes we found of the generalization error bound under different adversarial attacks and different training states, we proposed an adaptive training method which can greatly improve the image manipulation ability of multi-scale GANs. The final experimental results show that our adaptive training method in this paper has greatly contributed to the improvement of the quality of the images generated by multi-scale GANs on several image manipulation tasks. In particular, for the image super-resolution restoration task, the multi-scale GAN model trained by the proposed method achieves a 100% reduction in natural image quality evaluator (NIQE) and a 60% reduction in root mean squared error (RMSE), which is better than many models trained on large-scale datasets.