Real-time Deep Video Deinterlacing

Fig. 1. (a) Input interlaced frames. (b) Deinterlaced results generated by SRCNN [4] re-trained with our dataset. (c) Blown-ups from (b) and (d) respectively. (d) Deinterlaced results generated by our method. The classical super-resolution method SRCNN reconstruct each frame based on a single field and has large information loss. It also follows the conventional translation-invariant assumption which does not hold for the deinterlacing problem. Therefore, it inevitably generates blurry edges and artifacts, especially around sharp boundaries. In contrast, our method can circumvent this issue and reconstruct frames with higher visual quality and reconstruction accuracy.