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

1. Problem Statement

Single-image super-resolution (SR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs. While recent hybrid CNN-Transformer architectures have achieved strong results, they face a critical bottleneck: scaling attention windows or convolution kernels to expand the receptive field significantly increases computational cost and memory usage.

• Limitations of Current Methods:
  • CNNs: Effective for local dependencies but struggle to capture global context without massive parameter counts.
  • Transformers: Excellent for global context but suffer from quadratic computational complexity ($O(N^2)$) regarding spatial resolution, making them inefficient for resource-constrained devices.
  • Linear Attention: Offers $O(N)$ complexity but often suffers from rank collapse, leading to low feature diversity and reduced representational capacity.
  • Existing Lightweight Models: Often sacrifice representational richness for efficiency or fail to capture long-range dependencies effectively.

The core challenge is achieving a superior trade-off between accuracy, efficiency, and scalability by expanding the effective receptive field without incurring heavy computational overhead.

2. Methodology: UCAN Architecture

The authors propose UCAN (Unified Convolutional Attention Network), a lightweight architecture that unifies convolution and attention mechanisms. The network follows a layered design:

1. Shallow Feature Extraction: A $3 \times 3$ convolution maps the LR input to shallow features.
2. Core Encoder (BERFG): The Broad Effective Receptive Field Group processes features through a "Sharing Block" and a "Receiving Block."
3. Reconstruction: Fused features are passed through a reconstruction module (convolution + pixel shuffle) to generate the HR output.

Key Technical Components

A. Hedgehog Attention (HgA)
To address the rank deficiency in linear attention, the authors introduce Hedgehog Attention.

• Mechanism: It utilizes a Hedgehog Feature Map (HFM) that concatenates $m$ symmetric exponential feature pairs: $[\exp(W^T X + b), \exp(-W^T X - b)]$.
• Benefit: Unlike standard ReLU or ELU, HFM preserves information from both positive and negative pre-projection directions. This prevents rank collapse, enhances feature diversity, and allows the model to capture richer, more expressive features without external rank recovery modules.

B. Semi-Sharing Mechanism (Sharing & Receiving Blocks)
To reduce computational redundancy while maintaining performance, UCAN employs a semi-sharing strategy within the Broad Effective Receptive Field Group (BERFG):

• Sharing Block (SB): Computes the full attention maps ($A_{qk}$ and $A_{map}$) using Shared Hybrid Attention (SHA).
• Receiving Block (RB): Reuses the pre-computed attention maps from the SB in its Received Hybrid Attention (RHA) modules.
• Result: This avoids recomputing expensive attention matrices for every layer, significantly lowering the computational burden while refreshing feature representations dynamically.

C. High Performance Attention (HPA) & Flash Attention

• Local Context: Uses a Convolutional MLP (ConvMLP) with a $7 \times 7$ kernel to capture local details without explicit QKV projections.
• Large Window Attention: Employs Flash Attention with a $32 \times 32$ window size. This enables exact attention computation with significantly lower memory usage and faster execution compared to standard self-attention, making large-window modeling feasible.

D. Dual Fusion Layer (DFL)
The DFL aggregates global spatial and channel information efficiently:

• Spatial Branch: Combines Hedgehog Attention with a lightweight depthwise convolution to model complex spatial dependencies.
• Channel Branch: Uses standard channel attention (softmax-based) along the channel dimension.
• Complexity: Both branches achieve linear complexity ($O(N)$) with respect to spatial resolution, ensuring scalability.

E. Large Kernel Distillation (LKD)
To preserve high-frequency structures and hierarchical textures without heavy computation:

• Strategy: Splits channels into a fine-grained subset (processed by a Triple Feature Extraction module) and a coarse subset (bypassed).
• Triple Feature Extraction: Consists of three parallel branches: Channel Attention, a local bottleneck ($1 \times 1 \to 3 \times 3 \to 1 \times 1$), and a hierarchical large-kernel branch using dilated convolutions.
• Benefit: Restricts heavy computation to a subset of channels while expanding the receptive field via dilation and knowledge distillation.

3. Key Contributions

1. Hedgehog Attention: A novel mechanism that enhances feature diversity in linear attention by increasing feature rank through symmetric exponential feature maps, eliminating the need for auxiliary rank recovery modules.
2. Unified Efficient Framework (UCAN): A model that successfully captures local and global dependencies across multiple receptive fields by integrating:
   • Flash Attention for efficient large-window processing.
   • Hedgehog Attention for robust global modeling under limited resources.
   • Multi-kernel convolution for essential local feature extraction.
3. Semi-Sharing & Distillation Architecture: A novel design that balances parameter sharing with task-specific specialization. It reduces parameters and computational cost while maintaining strong reconstruction quality.

4. Experimental Results

The authors evaluated UCAN on five standard benchmarks: Set5, Set14, BSDS100, Urban100, and Manga109.

• Performance vs. Efficiency:
  • Manga109 (4×): The UCAN-L variant achieves 31.63 dB PSNR with only 48.4G MACs, surpassing recent lightweight models like MambaIRV2 (by 0.39 dB) with up to 36% lower computational cost.
  • BSDS100: UCAN achieves 27.79 dB, outperforming methods with significantly larger models.
  • Urban100 (2×): UCAN surpasses OmniSR by 0.12 dB using 11% fewer parameters and 24% fewer FLOPs.
• Comparisons: UCAN consistently outperforms state-of-the-art CNN-based, Transformer-based, and State Space Model (SSM) based lightweight SR methods (e.g., SwinIR-light, OmniSR, ATD-light, MambaIR-light).
• Qualitative Analysis:
  • Effective Receptive Field (ERF): Visualizations show UCAN has significantly darker and more extended regions of influence compared to competitors, indicating better global context capture.
  • Local Attribution Maps (LAM): UCAN demonstrates superior ability to aggregate information from broader contexts, including repetitive patterns, compared to models with fixed small windows.

5. Significance

UCAN represents a significant advancement in lightweight image super-resolution. By unifying convolutional efficiency with attention-based global modeling, it solves the long-standing trade-off between receptive field size and computational cost.

• Practical Deployment: Its low parameter count and reduced MACs make it highly suitable for deployment on resource-constrained devices (e.g., mobile phones, edge devices).
• Architectural Innovation: The introduction of Hedgehog Attention and the semi-sharing mechanism provides a new blueprint for designing efficient deep learning models that do not sacrifice representational power for speed.
• State-of-the-Art: It establishes a new benchmark for high-fidelity image restoration, proving that expansive receptive fields can be achieved without the prohibitive costs of full Transformer architectures.