Abstract
Image Super-Resolution (SR) has traditionally relied on increasingly deep and wide convolutional neural networks, leading to high computational costs and a neglect of spatial relationships. This paper introduces the Lightweight Non-Local Network (LNLN), which utilizes depthwise separable convolutions and non-local modules to capture global spatial context efficiently. LNLN achieves state-of-the-art results while reducing model size from over 55MB to just 8.72MB.
Introduction
Reconstructing high-resolution (HR) images from low-resolution (LR) inputs is an ill-posed problem. While deep learning has significantly improved SR performance, modern architectures like RCAN or EDSR are often too large for edge devices. Furthermore, many models rely on channel attention, which can overlook critical spatial dependencies, resulting in blurred textures. LNLN addresses these challenges by balancing representational power with extreme parameter efficiency.
Architecture and Methodology
The LNLN architecture is designed around the Lightweight Residual Group (LRG), which replaces standard convolutions with more efficient alternatives.
Key Components:
- Non-Local Modules: Unlike standard convolutions that operate on local neighborhoods, non-local modules calculate dependencies between every pixel pair in a feature map. This allows the network to capture long-range spatial information essential for reconstructing repetitive patterns.
- Depthwise Separable Convolutions (DSC): By splitting standard convolutions into depthwise and pointwise stages, LNLN drastically reduces computational complexity.
- Regional Grids: To maintain efficiency, non-local operations are performed within regional grids, providing a large receptive field without the quadratic cost of full-image attention.
Experimental Results
LNLN was evaluated on standard benchmarks including Set5, Set14, BSD100, Urban100, and Manga109.
Performance and Efficiency
The model demonstrates superior performance on the Urban100 dataset, achieving a PSNR of 26.96 dB at scale factor 4, surpassing the much larger RCAN (26.82 dB). Most notably, LNLN uses only 2.04 million parameters—roughly 14% of the parameter count of previous state-of-the-art modules. This efficiency results in a storage requirement of only 8.72 MB, compared to 55.4 MB for DRCA.
Conclusion
The Lightweight Non-Local Network proves that high-quality super-resolution does not require massive model sizes. By prioritizing spatial context through non-local modules and optimizing convolution types, LNLN offers a viable solution for high-performance SR on resource-constrained devices. Future work will explore semi-supervised learning to further streamline the training process.
Summary updated