Embedded Block Residual Network: A Recursive Restoration Model for Single-Image Super-Resolution

Yajun Qiu, Ruxin Wang, Dapeng Tao, Jun ChengICCV 2019 · DOI

Introduction

Single-Image Super-Resolution (SISR) is a classic ill-posed problem: reconstructing a high-resolution (HR) image from a single low-resolution (LR) input. While deep learning has significantly advanced SISR, many existing architectures suffer from a "one-size-fits-all" approach, applying the same network depth to all image regions regardless of their complexity.

Images naturally contain a variety of frequency bands. Low-frequency components, such as smooth gradients and simple shapes, require relatively little computational effort to restore. In contrast, high-frequency components, like fine textures and sharp edges, are much more complex. Standard deep networks often overfit simple textures, producing "wavy" artifacts, or underfit complex ones, resulting in blurriness. The Embedded Block Residual Network (EBRN) addresses this by matching model complexity to image frequency.

The EBRN Architecture

The core of EBRN is a recursive architecture designed for frequency-aware restoration. It moves beyond simple stacking of layers to an "embedded" structure where shallower modules handle low-frequency information and deeper modules focus on high-frequency details.

The EBRN architecture consists of feature extraction, embedded block residual learning, and reconstruction stages, mapping network depth to image frequency.

Block Residual Module (BRM)

Each BRM contains two distinct data flows:

  1. Super-resolution Flow: Utilizes a deconvolutional layer followed by three convolutional layers to restore structures at a specific frequency level.
  2. Back-projection Flow: This flow calculates the "residual"—the information the current module failed to recover—and passes it to the next, deeper BRM.

Embedded and Recursive Design

Multiple BRMs are linked such that the first module handles the lowest frequency components. Its residual (the high-frequency error) becomes the input for the subsequent module. This ensures that deeper, more complex sub-networks are dedicated solely to the most challenging details. To integrate these components, EBRN employs a recursive fusion technique where the output of a deeper module is summed with its predecessor and refined through a convolutional layer, stabilizing gradient flow and smoothing the final reconstruction.

Key Results and Performance

EBRN was evaluated on standard benchmarks including Set5, Set14, B100, Urban100, and Manga109. Quantitatively, it achieved state-of-the-art PSNR and SSIM scores, often outperforming much larger models like RCAN and EDSR while maintaining a smaller parameter footprint. For instance, on Set5 (×4), EBRN reached a PSNR of 32.79 dB.

Qualitative comparison on the Barbara dataset demonstrates EBRN's superior ability to restore repetitive book textures without the aliasing artifacts seen in other models.

Qualitatively, EBRN excels at restoring repetitive textures. As seen in the "Barbara" dataset, EBRN avoids the "checkerboard" or "aliasing" artifacts that plague other deep models when processing dense, high-frequency patterns.

Conclusion

The Embedded Block Residual Network demonstrates that frequency-aware modeling is a powerful paradigm for SISR. By aligning network depth with the complexity of image features, EBRN achieves superior restoration quality and parameter efficiency. This "coarse-to-fine" approach, inspired by classical wavelet analysis, provides a robust framework for future image restoration tasks such as denoising and deblurring.

Summary updated

BibTeX
@inproceedings{qiu2019embedded,
  author = {Yajun Qiu and Ruxin Wang and Dapeng Tao and Jun Cheng},
  title = {{Embedded Block Residual Network: A Recursive Restoration Model for Single-Image Super-Resolution}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2019},
  doi = {10.1109/iccv.2019.00428},
}
Figure