An Enhanced Video Super Resolution System Using Group-Based Optimized Filter-Set with Shallow Convolutional Neural Network

Scaling up video resolution has conventionally been achieved via linear interpolation, however this method occasionally introduces blurring to the output. Super-resolution (SR), an approach to preserve image quality in enlarged still images, has been exploited as a substitute for linear interpolation, however, the output at times exhibits image qualities worse than what linear interpolation produces primarily because the initial goal of SR is preservation of image quality when a still image is enlarged. In this context, this paper proposes a fast-performance adaptive system for scaling-up other resolutions like X2 using X3 model or X3 using X2 model by (1) first grouping frames that would use similar filter sets (2) then conducting fine-tuning of shallow CNN for SR on each frame group. Filter sets fine-tuned for each group resulted in significantly improved PSNR over either linear interpolation or conventional SR in our experiment. In the fine-tuning stage for each group, 0.5K to 2.5K iterations were sufficient to improve PSNR by 10%. By fine-tuning instead of performing full training, the number of sufficient iterations was reduced from 3,000K to mere 0.5K to 2.5K.