Generative Adversarial Network-Based Frame Interpolation with Multi-Perspective Discrimination

Video frame interpolation plays an important role in video applications. This paper presents a new deep learning-based frame interpolation model under the generative adversarial networks (GANs) framework. The devised generator cascades three multi-scale sub-modules to capture coarse-to-fine visual characteristics. Furthermore, the generator is trained with an adequate loss function consisting of two adversarial and four reconstruction losses. Additional to the conventional discriminator that derives the spatial discrepancy, a new discriminator is employed to strengthen the temporal consistency among adjacent frames. The proposed method not only inherits the merit of GANs in processing speed but also produces high-quality video frames. Experimental results show that the proposed method achieves the most favorable image quality in PSNR and SSIM with a considerably short runtime compared to other state-of-the-art frame interpolation approaches.