Abstract
Single Image Super-Resolution (SISR) is a challenging task aimed at recovering high-resolution (HR) images from their low-resolution (LR) counterparts. Traditional convolutional neural networks (CNNs) often fail to capture long-range pixel dependencies and complex, high-order statistical information necessary for reconstructing fine textures. This paper introduces the Mixed High-Order Non-local Attention Network (MHNAN), which integrates a novel Mixed High-Order Attention (MHA) module and a Non-local Enhanced Group (NEG) backbone. By modeling higher-order feature inter-dependencies and global context, MHNAN achieves state-of-the-art performance across multiple benchmark datasets.
Introduction
The goal of SISR is to reconstruct a visually pleasing HR image from a single LR input. Despite significant progress driven by deep learning, two primary challenges remain: the limited receptive field of standard convolutions and the simplicity of existing first-order attention mechanisms. Standard CNNs focus on local neighborhoods, neglecting global structural information. Furthermore, common attention modules (e.g., channel attention) typically model linear relationships, which are insufficient for capturing the intricate textures found in natural images.
Proposed Method
The MHNAN architecture is designed to address these limitations through two key components: the Mixed High-Order Attention (MHA) module and the Non-local Enhanced Group (NEG).
Overall Architecture
The network follows a modular design consisting of shallow feature extraction, a deep feature extraction backbone (the NEG), and an upscaling reconstruction module.
Mixed High-Order Attention (MHA)
The MHA module moves beyond standard first-order attention by calculating higher-order representations (up to the 4th order). This allows the network to focus on discriminative, high-frequency details and complex textures that are often lost in lower-order processing. The MHA modules are strategically placed before and after the main NEG backbone to refine features at different stages.
Non-local Enhanced Group (NEG)
The NEG serves as the core of the network, comprising several Non-local Residual Channel Attention Groups (NRCAG). Each NRCAG integrates non-local blocks to capture long-range correlations across the entire image, ensuring structural consistency in the reconstructed output.
Experiments and Results
The MHNAN was trained on the DIV2K dataset and evaluated on five standard benchmarks: Set5, Set14, BSD100, Urban100, and Manga109.
Quantitative Performance
MHNAN consistently outperformed 14 state-of-the-art methods, including RCAN and EDSR, in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). For instance, on the BSD100 dataset at a 4x scaling factor, MHNAN achieved a PSNR of 27.81 dB.
Qualitative Analysis
Visual comparisons demonstrate MHNAN's superior ability to reconstruct repetitive patterns and fine details, such as architectural features in the Urban100 dataset, without the aliasing artifacts common in other models. Ablation studies confirmed that 4th-order attention provides an optimal balance between reconstruction quality and computational efficiency.
Conclusion
The Mixed High-Order Non-local Attention Network (MHNAN) represents a significant advancement in SISR by leveraging high-order feature statistics and global context. The combination of the MHA module and the NEG backbone allows for the reconstruction of high-fidelity images with superior texture and structural integrity. Future work will focus on optimizing these high-order mechanisms for real-time and mobile applications.
Summary updated