Deep Learning based Real-Time Image Upscaling for Limited Data Rate and Prevalent Resources

Video streaming platforms have attracted substantial attention. The demand of maximizing image quality at the user end and in parallel, minimizing the amount of data transmission or date rate, is significantly increasing. On the other hand, compression algorithms have almost reached the upper bound; therefore, the only way to significantly decrease the amount of data transmission or data rate would be to decrease the image resolution itself on the transmitter (Tx) and to upscale the image on the receiver (Rx) or user side. The difficulty of this lies in the fact that users have a variety of resources available to perform upscaling, and normal upscaling does not yield sufficiently high-quality images. Deep learning has enabled the advent of image super-resolution techniques; however, such solutions require enormous time at the user end, which is inappropriate for real-time applications. With the aforementioned motivation, this paper proposes a deep learning network that can be run in real-time on the prevalent resources at the user end, e.g., a single core of the 2 GHz CPU, while providing the better quality than conventional methods of interpolation.