Partial Filter-Sharing: Improved Parameter-sharing Method for Single Image Super-Resolution Networks

Karam Park, Nam Ik ChoWACV 2025 · DOI

Abstract

Single Image Super-Resolution (SISR) often faces a trade-off between model capacity and parameter efficiency. Traditional parameter-sharing methods reduce model size but limit representational power by forcing all layers to use identical filters. This paper introduces Partial Filter-Sharing (PS), a novel framework that shares filter segments and uses small, layer-specific coefficient matrices to reconstruct unique filters for each layer. This approach significantly improves image reconstruction quality with minimal computational overhead. Experiments on standard benchmarks demonstrate that PS-enhanced models consistently outperform state-of-the-art recursive networks like CARN and SRFBN.

Introduction

Single Image Super-Resolution (SISR)—the process of reconstructing a high-resolution image from a single low-resolution input—is a fundamental task in computer vision. While deep learning has pushed the boundaries of SISR performance, it has often come at the cost of massive parameter counts, making high-performing models difficult to deploy on mobile or real-time systems.

To address this, researchers have turned to parameter sharing, such as in recursive or recurrent networks, where the same convolutional filters are reused across multiple layers. However, this creates a "capacity bottleneck": because every layer is forced to use the exact same weights, the network cannot easily learn the distinct, hierarchical features required at different stages of the upscaling process.

The Partial Filter-Sharing (PS) Framework

The proposed Partial Filter-Sharing (PS) method aims to restore this lost flexibility without sacrificing efficiency. Instead of sharing a single, static weight matrix across layers, the PS framework decomposes the filters into two components:

  1. Partial Filters ($P$): A shared pool of filter "fragments" or segments that act as a global vocabulary.
  2. Coefficient Matrices ($W_{c,i}$): A small set of unique weights for each layer $i$ that dictate how to combine the partial filters.

To generate the actual filter for a specific layer, the network performs a matrix multiplication followed by a reshaping operation:
$W_i = Reshape(P \times W_{c,i})$

This allows each layer to "reconstruct" its own unique filter from the shared partial filters. It provides a middle ground between full parameter sharing (high efficiency, low capacity) and no parameter sharing (low efficiency, high capacity).

Comparison of traditional parameter-sharing (a) and the proposed Partial Filter-Sharing method (b).

Experimental Evaluation

The authors trained their models on the DIV2K dataset and evaluated performance on four standard benchmarks: Set5, Set14, B100, and Urban100. The primary metrics used were PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity), alongside Multi-Adds to measure computational complexity.

The PS framework was integrated into several existing architectures, including CARN (Cascading Residual Network), DRRN (Deep Recursive Residual Network), and SRFBN (Super-Resolution Feedback Network).

Results and Discussion

The experimental results highlight several key advantages of the PS method:

  • Performance Gains: PS-enhanced models consistently outperformed their baseline counterparts. For example, CARN-PS achieved higher PSNR than the original CARN while actually using 10% fewer parameters.
  • Efficiency: The computational overhead introduced by the filter reconstruction step is negligible. In most cases, inference time increased by less than 2%. For instance, SRFBN-S-PS added only 0.06G Multi-Adds to a baseline of 852.9G.
  • Representational Diversity: By allowing layer-specific mixing, PS enables the network to capture more complex correlations. The authors found that the way filters are "sliced" (spatially vs. channel-wise) is crucial, recommending a single-direction split for optimal learning.
PSNR vs. Parameter count for various SISR models and their PS-enhanced versions. Red dots indicate PS-enhanced versions, showing improved performance at similar or lower parameter counts.

Conclusion

Partial Filter-Sharing (PS) provides a powerful and lightweight solution to the capacity bottleneck in recursive SISR networks. By reconstructing unique filters from a shared pool of partial filters, the method achieves a superior balance between performance and parameter efficiency. Future work may extend this framework to other domains, such as video processing or object detection, where efficient yet expressive models are highly desirable.

Summary updated

BibTeX
@inproceedings{park2025partial,
  author = {Karam Park and Nam Ik Cho},
  title = {{Partial Filter-Sharing: Improved Parameter-sharing Method for Single Image Super-Resolution Networks}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
  doi = {10.1109/wacv61041.2025.00263},
}
Figure