dcscn-super-resolution

Why we included this project

DCSCN is a compact TensorFlow implementation of a single-image super-resolution architecture that pairs a deep CNN with residual blocks, skip connections, and network-in-network style 1x1 convolutions. Beyond the original paper, the code adds pixel shuffler upsampling, optional transposed convolutions, self-ensemble inference, and depthwise separable convolutions, so it reads as a small tour of several classic SISR techniques in one place. That makes it handy for teams building or benchmarking their own upscaling pipeline who want to see how these pieces fit together and what they do to PSNR without adopting a large framework. Researchers and students reproducing the paper's results will also find it a straightforward base for trying architectural variations.