Realizing Real-Time Deep Learning-Based Super-Resolution Applications On Integrated Gpus

With recent advances in deep convolutional neural networks (CNN), deep learning has brought significant quality improvement and flexibility on single image super resolution (SR). In this paper, we describe how CNN based SR can be accelerated on integrated GPUs. To this end, we employ a CNN model from an existing single image SR approach, and develop the model within a well-known deep learning framework with OpenCL (TM) support. We also introduce a multi-tile approach in which we divide a large input into smaller tiles to generate SR for better utilization of memory bandwidth and to overcome size constraints posed by certain frameworks and devices thereby improving performance. This contributes to extending single image SR to video SR as well where video frames are considered as a group of multiple tiles. We prove that our approach is useful to resolve memory issues in generating ultra-high SR and to speed-up CNN based SR up to 44fps to generate various sizes of SR without quality impact.