Accelerating the Super-Resolution Convolutional Neural Network

Building on the successful SRCNN model for image super-resolution, this paper targets the high computational cost that hindered practical, real-time (24 fps) usage. The authors propose a compact hourglass-shaped CNN structure with three main redesigns: introducing a deconvolution layer at the end of the network so the mapping is learned directly from the original low-resolution image without interpolation; reformulating the mapping layer by shrinking the input feature dimension before mapping and expanding it back afterward; and adopting smaller filter sizes but more mapping layers.

The resulting model achieves a speedup of more than 40 times over SRCNN while delivering even superior restoration quality. The authors further present parameter settings that reach real-time performance on a generic CPU while maintaining good quality, and propose a corresponding transfer strategy for fast training and testing across different upscaling factors. This made high-quality super-resolution far more practical for real-world, latency-sensitive applications.