Video Super-Resolution using Multi-Frames Fusion and Perceptual Loss

Abstract

Video super-resolution (VSR) has become more and more important in vision perception. Currently, most of video super-resolution methods used optical flow estimation by explicit motion compensation. We proposed a new network using multi-frames as input without motion compensation. In order to improve the performance of the reconstructed high-resolution (HR) frame, perceptual loss function is used to evaluate the deviation between the prediction and ground truth. Perceptual loss characteristic is able to improve the quality of image reconstruction on the vision, but with a low Peak Signal to Noise Ratio (PSNR). Experimental results show that compared with other methods, our model can recover more details in high-resolution image, although it does not get higher PSNR value.