Skip to content

UCAN: Unified Convolutional Attention Network for Expansive Receptive Fields in Lightweight Super-Resolution

Conference: CVPR 2026
arXiv: 2603.11680
Code: https://github.com/hokiyoshi/UCAN
Area: Image Restoration / Lightweight Super-Resolution
Keywords: Lightweight super-resolution, Hedgehog attention, large-kernel distillation, receptive field expansion, parameter sharing

TL;DR

UCAN is a lightweight super-resolution network that unifies convolutional and attention mechanisms to efficiently expand the effective receptive field. It addresses the rank collapse issue of linear attention via Hedgehog attention, introduces a large-kernel distillation module and a semi-shared parameter strategy, and achieves 31.63 dB PSNR on Manga109 (×4) with only 48.4G MACs.

Background & Motivation

  1. Background: Lightweight SR methods primarily improve performance by expanding the effective receptive field. Transformer-based approaches are effective but significantly increase computational cost when enlarging attention windows or convolution kernels.
  2. Limitations of Prior Work: Global attention methods such as Grid Attention and Mamba still suffer from efficiency issues. Linear attention achieves \(O(N)\) complexity but exhibits rank collapse, leading to insufficient feature diversity. Parameter sharing and distillation strategies may homogenize feature maps.
  3. Key Challenge: An inherent tension exists between expanding the receptive field and maintaining a lightweight design, as well as a trade-off between efficiency and representational richness.
  4. Goal: To simultaneously model local textures and global dependencies under lightweight constraints.
  5. Key Insight: Hedgehog feature mapping is employed to address rank collapse in linear attention, and Flash Attention is used to enable efficient computation over large attention windows.
  6. Core Idea: Multi-level fusion — Flash Attention handles large-window local interactions, Hedgehog Attention captures global dependencies, and large-kernel distillation convolutions model spatial structure.

Method

Overall Architecture

LR input → 3×3 convolution for shallow features → Broad Effective Receptive Field Group (BERFG, comprising shared blocks and receiving blocks) → residual connection → PixelShuffle reconstruction. BERFG contains High-Performance Attention (HPA), Hybrid Attention (SHA/RHA), and Large-Kernel Distillation (LKD).

Key Designs

  1. Hedgehog Attention:

    • Function: Maintains high-rank feature representations under linear complexity.
    • Mechanism: Replaces simple mappings such as ReLU/ELU with a Hedgehog Feature Mapping (HFM). HFM concatenates \(m\) pairs of symmetric exponential features: \(\phi_H(X) = [\exp(W^\top X + b_1), ..., \exp(-W^\top X - b_m)]\). Symmetric pairing preserves information in both positive and negative directions, avoiding the information loss caused by ReLU discarding negative values and the extreme variations introduced by ELU+1. Empirically, linear attention with HFM recovers rank to 46 out of 64 (full rank), far exceeding ReLU/ELU.
    • Design Motivation: Low-rank output matrices in linear attention compress features into few directions, resulting in insufficient representational diversity. The trainable MLP-style structure of HFM offers greater flexibility than fixed mappings.

  2. Semi-Shared Mechanism:

    • Function: Maintains representational updates within parameter sharing.
    • Mechanism: BERFG is divided into Shared Blocks (SB) and Receiving Blocks (RB). The shared hybrid attention in SB computes full attention and caches \(A_{qk}^{(a)}, A_{map}^{(a)}\). The receiving hybrid attention in RB directly reuses the softmax attention maps from SB, while the dynamic feature mappings (\(\phi(Q), \phi(K)\)) of Hedgehog attention are independently recomputed at each layer.
    • Design Motivation: Full sharing leads to representational homogenization. Semi-sharing shares the window attention component (reducing computation) while independently updating the global attention component (preserving diversity).

  3. Large-Kernel Distillation Module (LKD):

    • Function: Expands spatial receptive field with low parameter overhead.
    • Mechanism: Channels are split into a fine-grained subset \(F_{fg}\) (\(\max(C/4, 16)\) channels) and a coarse-grained subset \(F_{cg}\). A three-branch feature extractor (TFE) is applied only to \(F_{fg}\): a channel attention branch, a 1×1→3×3→1×1 bottleneck local branch, and a hierarchical large-kernel branch using depthwise separable and dilated convolutions. \(F_{cg}\) is passed through directly.
    • Design Motivation: Confining heavy computation to a small subset of channels proportionally reduces overall cost, while the large-kernel path efficiently expands the receptive field via dilation and depthwise separable convolutions.

Loss & Training

L1 reconstruction loss + LDL loss + Wavelet loss. Adam optimizer (\(\beta_1=0.9, \beta_2=0.99\)), 64×64 crop, batch size 16. Trained on 2 × RTX 3090. ×2 scale trained from scratch for 800K iterations; ×3/×4 scales fine-tuned from ×2 for 400K iterations.

Key Experimental Results

Main Results

Method Manga109 4× PSNR Params MACs
UCAN-L 31.63 902K 48.4G
MambaIRV2-light 31.24 790K 75.6G
ATD-light 31.48 769K 100.1G
ESC 31.54 968K 149.2G
RCAN 31.22 15592K 917.6G

Ablation Study

Configuration Set5 PSNR Urban100 PSNR Notes
w/o HPA 38.27 32.90 Missing large-window local attention
HPA 16×16 window 38.32 33.04 Default 32×32 is superior
ReLU mapping 38.33 33.16 Low rank
Hedgehog mapping 38.34 33.22 High rank, +0.06 dB
Full sharing 38.29 32.89 Representational homogenization
Semi-sharing 38.34 33.22 Maintained diversity, +0.33 dB

Key Findings

  • UCAN outperforms MambaIRV2 by 0.39 dB on Manga109 (×4) while reducing MACs by 36%.
  • Hedgehog feature mapping recovers rank to 46/64; ReLU and ELU reach only ~20 and ~30, respectively.
  • ERF visualization demonstrates that UCAN's effective receptive field coverage is substantially larger than that of MambaIR/MambaIRv2.
  • LAM analysis shows that UCAN aggregates repeated patterns and similar structures from a broader context.

Highlights & Insights

  • Hedgehog Attention Resolves Rank Collapse: Symmetric exponential feature mapping restores the rank of linear attention, directly improving representational diversity.
  • Multi-Level Receptive Field Fusion: Flash Attention (32×32 local) + Hedgehog Attention (global) + Large-Kernel Distillation (spatial structure) form a complementary three-way design.
  • Extreme Efficiency: With only 705K parameters and 38.1G MACs, UCAN achieves performance comparable to RCAN (15.6M parameters, 918G MACs).

Limitations & Future Work

  • Flash Attention relies on specific CUDA implementations and may not be available on certain hardware platforms.
  • The number of feature pairs \(m\) in the Hedgehog feature mapping requires tuning.
  • Validation is limited to SR tasks; generalizability to other image restoration tasks remains to be verified.
  • vs. OmniSR: OmniSR expands the receptive field via Grid Attention but with limited efficiency; UCAN achieves greater efficiency.
  • vs. MambaIRv2: MambaIRv2 combines Swin Transformer with SSM; UCAN replaces SSM with Hedgehog linear attention.
  • vs. ATD-light: ATD employs an adaptive token dictionary, while UCAN uses large-kernel distillation and Hedgehog attention, achieving lower MACs.

Rating

  • Novelty: ⭐⭐⭐⭐ First application of Hedgehog attention in SR with rank recovery analysis.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Five benchmarks, three scales, ERF/LAM analysis, and detailed ablations.
  • Writing Quality: ⭐⭐⭐⭐ Clear structure with in-depth analysis of attention mechanisms.
  • Value: ⭐⭐⭐⭐ Establishes a new SOTA direction for lightweight SR.