Super‐resolution with adversarial loss on the feature maps of the generated high‐resolution image

Abstract

Abstract Recent studies on image super‐resolution make use of Generative Adversarial Networks to generate the high‐resolution image counterpart of the low‐resolution input. However, while being able to generate sharp high‐resolution images, Generative Adversarial Networks based super‐resolution methods often fail to produce good results when tested on images having different degradation as the low‐resolution images used in the training. Some recent works have tried to mitigate this failure by introducing a degradation network that can replicate the noise of real‐world low‐resolution images. However, even these methods can produce poor results if a real‐world test image differs much from the real‐world images in the training data set. This paper proposes the use of adversarial losses on the feature maps extracted by a pre‐trained network with the generated high‐resolution image as input. This is in contrast to all other Generative Adversarial Networks‐based super‐resolution methods that directly apply the adversarial loss to the generated high‐resolution image. The rationale behind this idea is illustrated, and experimental results confirm that high‐resolution images generated by the proposed method achieve better results in both quantitative and qualitative evaluations than methods that directly apply adversarial losses to generated high‐resolution images.