Edge-Based Loss Function for Single Image Super-Resolution

Abstract

In recent years, convolutional neural networks have shown state-of-the-art performance on the task of single-image super-resolution. Although these proposed networks have shown high-quality reconstruction results, the use of the mean-squared error (MSE) loss function for training tends to produce images that are overly smooth and blurry. The MSE does not consider image structures that are often important for achieving high human-perceived image quality. We propose a novel edge-based loss function to improve super-resolution resconstruction of images. Our loss function directly optimizes the edge pixels of the reconstructed image, thus driving the trained network to produce high-quality salient edges and thus sharper images. Extensive quantitative and qualitative results show that our proposed loss function significantly outperforms the MSE.