Edge-Aware Normalized Attention for Efficient and Detail-Preserving Single Image Super-Resolution

2025-09-17

AI Summary

Single-image super-resolution (SISR) suffers from structural distortions and artifacts due to the severe ill-posedness of high-frequency detail recovery. To address this, we propose an Edge-aware Normalization Attention Mechanism (ENAM), which embeds edge priors into a lightweight residual network. ENAM jointly encodes edge features and intermediate activations to generate adaptive modulation maps, thereby enhancing structurally salient regions while suppressing artifacts. The method employs end-to-end adversarial training, integrating pixel-wise, perceptual, and adversarial losses in a multi-branch framework. Extensive experiments on standard benchmarks demonstrate that our approach significantly outperforms SRGAN, ESRGAN, and existing edge-attention methods in edge fidelity and visual realism. Despite comparable parameter count, our model features a more compact architecture and improved training stability, achieving a favorable balance between reconstruction quality and computational efficiency.

Technology Category

• Computer Vision: Adversarial Attacks & Robustness
• Machine Learning: Learning on the Edge & Model Compression
• Natural Language Processing: Generation

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Abstract

Single-image super-resolution (SISR) remains highly ill-posed because recovering structurally faithful high-frequency content from a single low-resolution observation is ambiguous. Existing edge-aware methods often attach edge priors or attention branches onto increasingly complex backbones, yet ad hoc fusion frequently introduces redundancy, unstable optimization, or limited structural gains. We address this gap with an edge-guided attention mechanism that derives an adaptive modulation map from jointly encoded edge features and intermediate feature activations, then applies it to normalize and reweight responses, selectively amplifying structurally salient regions while suppressing spurious textures. In parallel, we integrate this mechanism into a lightweight residual design trained under a composite objective combining pixel-wise, perceptual, and adversarial terms to balance fidelity, perceptual realism, and training stability. Extensive experiments on standard SISR benchmarks demonstrate consistent improvements in structural sharpness and perceptual quality over SRGAN, ESRGAN, and prior edge-attention baselines at comparable model complexity. The proposed formulation provides (i) a parameter-efficient path to inject edge priors, (ii) stabilized adversarial refinement through a tailored multiterm loss, and (iii) enhanced edge fidelity without resorting to deeper or heavily overparameterized architectures. These results highlight the effectiveness of principled edge-conditioned modulation for advancing perceptual super-resolution.

Problem

• Addresses ambiguous high-frequency recovery in single-image super-resolution
• Reduces redundancy and instability in edge-aware SISR methods
• Enhances structural sharpness without increasing model complexity

Innovation

• Edge-guided attention mechanism for adaptive modulation
• Lightweight residual design with composite objective training
• Parameter-efficient edge prior injection without deep architectures