Source-linked AI summary
Lightweight Image Super-Resolution with Adaptive Weighted Learning Network
Chaofeng Wang, Zheng Li, Jun Shi
TL;DR
Single-image super-resolution remains an ill-posed reconstruction problem, and computational costs limit the real-world use of many deep SR models. AWSRN combines adaptive weighted residual learning with multi-scale reconstruction; the paper reports better performance than state-of-the-art algorithms without increasing parameters or computational overhead.
Problem
SISR is an ill-posed task of reconstructing a high-resolution image from a low-resolution one, and deep SR models face computational burdens that limit real-world applications.
Source: §1 · pp. 1–2Method
AWSRN combines adaptive weighted residual units and local residual fusion in an LFB with an adaptive weighted multi-scale reconstruction module.
Source: §1 · p. 2Results
The authors report better performance than state-of-the-art algorithms without increasing parameters or computational overhead.
Source: §5 · p. 8Takeaways & Limitations
AWSRN is presented as a lightweight and efficient super-resolution network for SISR.
Source: §5 · p. 8Higher-performing EDSR, D-DBPN, RDN, and RCAN models have more than 10M parameters; AWSRN's x8 model has 2348K parameters, under 5% of EDSR's.
Source: §4.4 · p. 6
Abstract
from arXiv · showhide
Deep learning has been successfully applied to the single-image super-resolution (SISR) task with great performance in recent years. However, most convolutional neural network based SR models require heavy computation, which limit their real-world applications. In this work, a lightweight SR network, named Adaptive Weighted Super-Resolution Network (AWSRN), is proposed for SISR to address this issue. A novel local fusion block (LFB) is designed in AWSRN for efficient residual learning, which consists of stacked adaptive weighted residual units (AWRU) and a local residual fusion unit (LRFU). Moreover, an adaptive weighted multi-scale (AWMS) module is proposed to make full use of features in reconstruction layer. AWMS consists of several different scale convolutions, and the redundancy scale branch can be removed according to the contribution of adaptive weights in AWMS for lightweight network. The experimental results on the commonly used datasets show that the proposed lightweight AWSRN achieves superior performance on x2, x3, x4, and x8 scale factors to state-of-the-art methods with similar parameters and computational overhead. Code is avaliable at: https://github.com/ChaofWang/AWSRN
1. Introduction
SISR reconstructs high-resolution images from low-resolution ones, but remains an ill-posed inverse problem; deep models also face computational costs that limit real-world use. AWSRN combines adaptive weighted residual learning and multi-scale reconstruction, and reports state-of-the-art performance with a favorable operations trade-off.
1. Introduction
- Deep SR networks can deliver strong performance but heavy computation and large model sizes limit their real-world applications.Lightweight design is motivated by this resource burden.Source: §1 · p. 1
- Existing lightweight strategies reduce parameters through shallow or recursive networks, while NAS designs use fewer operations but have performance constrained by their search space and strategy.The introduction notes that some shallow and recursive approaches still have many Multi-Adds.Source: §1 · p. 1
- Residual-learning networks can face exploding gradients, while single-scale reconstruction may underuse nonlinear-mapping features and multi-scale reconstruction adds parameters and operations.Residual scaling uses fixed weights to limit gradient flow, but does not guarantee better performance.Source: §1 · pp. 1–2
- AWSRN uses an LFB of adaptive weighted residual units and a local residual fusion unit to support information and gradient flow and fuse multi-level residual information.The adaptive weights are described as helping information and gradient flow more efficiently.Source: §1 · p. 2
- Its AWMS reconstruction module uses nonlinear-mapping features to improve reconstruction quality and removes redundant scale branches based on their weighted contribution to reduce parameters.The network is presented as a feature-extraction module, nonlinear-mapping module, and AWMS reconstruction module.Source: §1 · p. 2
- AWSRN reports state-of-the-art performance on Set5 at x2 and a favorable trade-off between reconstruction performance and operations.The comparison visual contrasts AWSRN variants with other lightweight networks; Multi-Adds represents operations.Source: Figure 1 · p. 1, §1 · p. 2
2. Related Work
Earlier lightweight SR work reduces model size or computational cost through shallower networks, parameter sharing, and efficient reconstruction, while residual learning and multi-scale reconstruction motivate AWSRN’s design choices.
2.1. Lightweight Super-Resolution Networks
- Lightweight SR approaches include shallow networks, recursive parameter sharing, cascading residuals, and NAS, but some still incur many operations [9].Deeper models such as MDSR and RCAN are also described as increasing parameters and operations, limiting real-world use.Source: §2.1 · p. 2
2.2. Efficient Residual Learning
- Residual learning underpins VDSR, SRResNet, EDSR, and WDSR, while wSE generates residual-unit weights at added parameter and computational cost [17, 18, 19,20, 21, 22].AWSRN’s AWRU instead learns two independent weights without an additional module or computational overhead.Source: §2.2 · p. 2
2.3. Upsampling Layer
- Single-scale reconstruction may underuse nonlinear features, whereas multi-scale reconstruction can improve quality at the cost of more parameters and computation.AWMS uses learned branch contributions to remove lower-contribution scales and reduce parameters without losing performance.Source: §2.3 · p. 3
3. Method
AWSRN extracts features from the low-resolution image, maps them through stacked local fusion blocks, then reconstructs through AWMS. Each local fusion block combines adaptive weighted residual units with fusion of multilevel features.
3.1. Basic Network Architecture
- AWSRN extracts low-resolution features, processes them through stacked LFBs, and sends the final LFB output to AWMS for reconstruction.Its global residual path uses the LR image through a convolution and subpixelshuffle layers.Source: §3.1 · p. 3
3.2. Local Fusion Block
- Each LFB stacks AWRUs and uses an LRFU bottleneck to fuse multilevel features and match shortcut dimensions.The LFB also combines the residual output with a bottleneck transform of earlier AWRU features.Source: §3.2 · p. 4
- AWRU applies two independently learned weights to the residual and identity branches, unlike wRU’s additional wSE weight-generation module.The basic residual unit is the WDSR wide-activate unit, which expands internal dimensions before ReLU while shrinking input/output dimensions.Source: §3.2 · p. 4
4. Experiments
The experiments evaluate AWSRN on standard super-resolution benchmarks and test its components through ablations. The reported comparisons show strong performance across parameter scales, while probing how adaptive residual weighting and multi-scale reconstruction contribute.
4.1. Datasets and Metrics
- DIV2K’s 800 bicubic-downsampled image pairs train the models; evaluation uses five benchmarks and PSNR and SSIM on the Y channel.The benchmarks are Set5, Set14, B100, Urban100, and Manga109.Source: §4.1 · p. 4
4.2. Implementation Details
- Training uses 48×48 LR patches, random rotations and flips, ADAM with L1 loss, and a learning rate halved every 2×10^5 iterations.Each minibatch draws 16 randomly cropped patches; the initial learning rate is 10^-3.Source: §4.2 · p. 4
4.3. Ablation Study
- In WDSR, Table 1 reports that both wRU and AWRU achieve higher PSNR values than Basic RU, and that AWRU has fewer parameters than wRU.Source: §4.3 · p. 4, Table 1 · p. 5
- AWRU residual and shortcut weights decrease with network depth, while shallow layers have larger weights, especially on the shortcut branch.The authors interpret the smaller weights in deeper layers as helping prevent gradient explosion.Source: §4.3 · p. 5, Figure 4 · p. 5
- AWSRN-B outperforms variants without AWRU or LRFU, and AWMS outperforms a single 3×3 reconstruction convolution on the tested backbones.The AWMS comparison averages PSNR across benchmark datasets at scale 2.Source: §4.3 · p. 5, Table 3 · p. 5
- The 3×3 branch captures low-frequency information and larger branches respond to high-frequency edges; 3×3 and 5×5 matter more, while removing 7×7 has no significant effect.The branch-removal test reports average PSNR on Set5 at scale 2.Source: §4.3 · pp. 5–6, Table 4 · p. 5, Figure 5 · p. 6
4.4. Comparison with State-of-the-art Methods
- AWSRN models achieve the best performance in each parameter scale in Table 5, which reports PSNR and SSIM across five datasets alongside parameters and Multi-Adds.Multi-Adds were calculated for a 1280×720 SR image at all scales.Source: §4.4 · p. 6, Table 5 · p. 7
- AWSRN-S beats similarly sized models on ×2 and ×3, while AWSRN-SD is best among the small models; CARN-M is slightly better than AWSRN-S on ×4.AWSRN-SD has fewer parameters and Multi-Adds than AWSRN-S; the ×4 comparison attributes CARN-M’s slight advantage to its deeper network.Source: §4.4 · p. 6
5. Conclusions
AWSRN is presented as a lightweight and efficient SISR network. Its local fusion block supports residual learning, while its adaptive weighted multi-scale reconstruction uses context and branch redundancy to reduce parameters.
5. Conclusions
- AWSRN combines adaptive weighted residual units and local residual fusion in LFBs with AWMS reconstruction, which can reduce parameters by identifying redundant scale branches.The authors attribute efficient information and gradient flow and integration to LFB, and context use and redundancy analysis to AWMS.Source: §5 · p. 8
- The authors state that their adaptive weighting methods can be applied to other SR models or visual tasks.They also express hope that the work can be applied to real-world applications such as live video.Source: §5 · p. 8
- The conclusion describes the models as achieving better performance than state-of-the-art algorithms without increased parameters or computational overhead.Source: §5 · p. 8