WaveMixSR-V2: Enhancing Super-resolution with Higher Efficiency (Student Abstract)

Pranav Jeevan, Neeraj Nixon, Amit SethiAAAI 2025 · DOI

TL;DR

This paper introduces WaveMixSR-V2, an improved single image super-resolution (SISR) model that achieves state-of-the-art performance with significantly higher efficiency. By adopting a multi-stage resolution doubling approach and replacing transposed convolutions with PixelShuffle operations, WaveMixSR-V2 reduces parameters, computation, and latency while boosting throughput.

What Problem Does This Paper Address?

Single image super-resolution (SISR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs, a crucial task in image processing. While recent transformer-based models have shown promise by capturing long-range dependencies, they often suffer from quadratic complexity, demanding substantial resources and large datasets. Previous efficient models like WaveMixSR used a single-stage upsampling, which could limit fine-tuning across scales. This paper addresses the need for SISR models that are both highly accurate and computationally efficient, particularly for higher upscaling factors (e.g., 4x), without requiring massive datasets.

Methods

WaveMixSR-V2 builds upon the WaveMixSR architecture by introducing two key improvements:

  • Multi-stage Design: Instead of a single upsampling step for higher resolution tasks (like 4x), WaveMixSR-V2 employs a series of "resolution-doubling" 2x SR blocks. This means a 4x SR task is achieved by two sequential 2x SR blocks, allowing the model to progressively refine details at each scale, leading to better performance and reduced resource consumption compared to a single, large upsampling operation.
  • PixelShuffle Operation: The transposed convolution layers within the WaveMix blocks, which are parameter-heavy and can introduce checkerboard artifacts, are replaced by a PixelShuffle operation followed by a standard convolution layer. PixelShuffle efficiently upsamples feature maps by rearranging pixels, significantly reducing the number of parameters and improving efficiency while producing smoother outputs.

Datasets

The paper primarily uses the BSD100 dataset for evaluation. It also mentions that the model requires less training data than transformer-based models and outperforms models pre-trained on ImageNet-1k.

Key Insights and Ideas

  • Progressive Upsampling is Key: The multi-stage design demonstrates that breaking down high-resolution upscaling into smaller, sequential steps (e.g., 2x then 2x for 4x) allows for more effective detail refinement and better overall performance.
  • PixelShuffle for Efficiency: Replacing transposed convolutions with PixelShuffle is a highly effective strategy for reducing model complexity and computational cost in SISR, without sacrificing output quality.
  • Efficiency Without Compromise: WaveMixSR-V2 shows that achieving state-of-the-art results in SISR does not necessitate massive parameter counts or high computational demands.

Key Results

WaveMixSR-V2 achieves state-of-the-art (SOTA) performance on the BSD100 dataset for 2x SR. Key quantitative results include:

  • Parameter Efficiency: WaveMixSR-V2 uses less than half the number of parameters compared to the previous SOTA WaveMixSR (0.7M vs 1.7M for 4x SR).
  • Computational Efficiency: It requires fewer computations and has lower latency than WaveMixSR.
  • Throughput: WaveMixSR-V2 exhibits significantly higher throughput (82.6 fps vs 53.7 fps for 2x SR) compared to WaveMixSR.
  • Accuracy: It surpasses previous state-of-the-art models like WaveMixSR, HAT, and SwinFIR in terms of PSNR and SSIM on the BSD100 dataset for 2x SR.

Limitations and Future Work

Not discussed in the paper.

Summary updated

BibTeX
@inproceedings{jeevan2025wavemixsr,
  author = {Pranav Jeevan and Neeraj Nixon and Amit Sethi},
  title = {{WaveMixSR-V2: Enhancing Super-resolution with Higher Efficiency (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35262},
}
Figure