Image Super-Resolution with Perceptual Quality Assessment Guidance

Zhiliang Ma, Yang Zhao, Yuan Chen, Xiaojuan Zhang

Abstract: Generative Adversarial Networks (GAN) have demonstrated the potential to recover realistic details for single image super-resolution (SISR). However, without considering perceptual metrics in optimization, existing perceptual SR methods could not show stable performance. To further improve the visual quality of super-resolution results, we propose a SISR method driven by image perceptual quality assessment. Through evaluating the generation results of GAN-based SISR and utilizing the evaluation results as the loss function, the generator can be optimized to obtain results with better perceptual quality. Specifically, we use a distorted dataset consisting of images generated by traditional and GAN-based image enhancement algorithms, and present an image perceptual quality assessment method which is highly consistent with human subjective evaluation. Then, we evaluate the generation results of GAN-based SISR and the evaluation result is utilized as the quality constraint to guide SISR model to obtain better subjective perception generation results. Although the proposed method is simple, it is effective and can be easily inserted into the existing SISR method. Extensive experiments show that the proposed method can produce visually pleasing results and achieve the state-of-the-art (SOTA) performance in perceptual metrics.