Perceptual Loss for Training Multi-Image Super-Resolution

Super-resolution (SR), i.e. reconstruction of high-resolution (HR) images from low-resolution (LR) samples, plays a vital role in remote sensing. Multi-image SR (MISR) methods leverage information fusion from a series of LR images acquired over time to reconstruct HR images. Deep neural networks form the foundation of state-of-the-art MISR techniques, and selecting an appropriate metric for training them is crucial to obtain reliable results. In this paper, we propose a novel approach for training MISR networks by incorporating the learned perceptual image patch similarity (LPIPS) metric as the loss function. Traditionally, SR methods rely on metrics based on pixel-wise similarity, which may not be optimal when LR and HR images differ significantly in histogram characteristics. By utilizing the LPIPS metric, which is consistent with human perception, the training process is directed towards better reconstruction of image details rather than strict pixel-wise compliance. Experimental results on the MuS2 dataset, consisting of LR images from Sentinel-2 and HR images from WorldView-2, demonstrate the effectiveness of our approach.