Adaptive Local Implicit Image Function for Arbitrary-Scale Super-Resolution

Hongwei Li, Tao Dai, Yiming Li, Xueyi Zou, Shu-Tao Xia

Abstract

Traditional single image super-resolution (SISR) models are often limited to fixed integer scales. While the Local Implicit Image Function (LIIF) introduced arbitrary-scale capabilities by representing images as continuous functions, it suffers from structural distortions and ringing artifacts at high magnification. We present Adaptive Local Implicit Image Function (A-LIIF), which employs an adaptive ensemble of multiple multi-layer perceptrons (MLPs) to better capture local geometric variations. By learning to weight different "expert" networks for specific image regions, A-LIIF significantly improves edge sharpness and reconstruction fidelity across arbitrary scales.

Introduction

Single Image Super-Resolution (SISR) aims to reconstruct a high-resolution (HR) image from a single low-resolution (LR) input. Most deep learning-based SISR methods are designed for specific upsampling factors (e.g., 2x, 4x). The recent Local Implicit Image Function (LIIF) addressed this by treating an image as a continuous function, allowing for arbitrary-scale upsampling. However, LIIF's use of a single global MLP for all pixel coordinates often fails to account for the diverse mathematical characteristics of edges, textures, and smooth regions, leading to "ringing" artifacts—ghostly lines near sharp edges—especially at high scales like 12x.

Methodology

A-LIIF introduces an adaptive mechanism to handle local image variations more effectively. The architecture comprises three core components:

  1. Encoder: A backbone network (such as EDSR) extracts deep feature maps from the LR input.
  2. Expansion Network: This module analyzes local features and target coordinates to generate $K$ adaptive scores, which serve as dynamic weights.
  3. Adaptive MLP Ensemble: Instead of a single MLP, A-LIIF utilizes a set of $K$ distinct MLPs. The final pixel color is computed as a weighted linear combination of the outputs from these $K$ networks, guided by the expansion network's scores.

This approach allows the model to specialize different MLPs for various image structures—such as sharp boundaries or smooth gradients—ensuring a more flexible and accurate reconstruction without the "one-size-fits-all" limitation of previous implicit functions.

Experimental Setup

The model was trained on the DIV2K dataset (800 training images). Evaluation was performed on standard benchmarks including Set5, Set14, B100, and Urban100. These datasets provide a variety of challenges, from human portraits to complex architectural patterns, allowing for a robust comparison of super-resolution algorithms.

Results and Discussion

A-LIIF demonstrates superior performance over the baseline LIIF and traditional fixed-scale models. Quantitatively, A-LIIF achieves higher PSNR values across all tested scales. For instance, on the DIV2K validation set at 12x scale, A-LIIF reached an average PSNR of 23.71 dB, compared to LIIF’s 23.17 dB.

Qualitatively, the adaptive ensemble effectively mitigates the ringing artifacts common in LIIF. As shown in the comparison below, A-LIIF preserves sharper edges and finer details in architectural elements, resulting in a more natural reconstruction.

Comparison of Ground Truth (a), LIIF (b), and A-LIIF (c) at 12x scale. A-LIIF shows reduced ringing and higher PSNR on this sample.

The model also exhibits strong generalization to extreme scales (up to 30x) not encountered during training, highlighting the robustness of the adaptive coordinate-based representation.

Conclusion

A-LIIF advances arbitrary-scale super-resolution by replacing rigid global functions with an adaptive ensemble of local experts. While this increases computational complexity, the gains in visual quality and structural integrity are substantial. Future research may explore more efficient backbones, such as Vision Transformers, and optimize the trade-off between the number of MLPs and inference speed.

Code available at: https://github.com/LeeHW-THU/A-LIIF

Summary updated

BibTeX
@inproceedings{li2022adaptive,
  author = {Hongwei Li and Tao Dai and Yiming Li and Xueyi Zou and Shu-Tao Xia},
  title = {{Adaptive Local Implicit Image Function for Arbitrary-Scale Super-Resolution}},
  booktitle = {ICIP},
  year = {2022},
  doi = {10.1109/icip46576.2022.9897382},
}

Loading related content...

Figure