EDSR

EDSR, or Enhanced Deep Residual Networks for Single Image Super-Resolution, is a convolutional neural network designed for single-image super-resolution (SISR). It learns to convert a low-resolution image into a higher-resolution version by mapping bicubic or other interpolations toward a high-resolution image. The model achieved state-of-the-art results on several standard SR benchmarks at the time of its introduction.

Architecturally, EDSR builds on deep residual learning but makes several key modifications. It removes batch normalization layers from residual blocks, allowing the network to learn more expressive mappings. The core body is a very deep stack of residual blocks, and the network employs a residual-in-residual structure with a residual scaling factor to stabilize training of the deeper model. The design emphasizes depth and feature reuse, enabling higher-quality reconstructions without heavy pre- or post-processing.

Upsampling is performed with a single upscaling module, typically using a sub-pixel convolution (PixelShuffle) layer to increase the spatial resolution to the target scale (for example, 2x or 4x). The network comprises a shallow feature extraction stage, followed by the deep residual backbone, and ends with an upsampling layer and a final reconstruction convolution.

Training typically uses the DIV2K dataset and proof-of-concept loss functions such as L1 or perceptual losses, with standard data augmentation and optimization practices. EDSR’s emphasis on depth, the removal of BN, and residual-in-residual design influenced subsequent advances in SR research, helping drive later architectures that push SR quality further.