Source-linked AI summary
Efficient Residual Dense Block Search for Image Super-Resolution
Dehua Song, Chang Xu, Xu Jia, Yiyi Chen, Chunjing Xu, Yunhe Wang
TL;DR
Deep-learning super-resolution faces computation and memory demands that hinder practical use, particularly on mobile devices. This paper searches efficient residual dense block networks using evolutionary architecture search and block credits. The authors report better performance than state-of-the-art methods with limited parameters and FLOPs.
Problem
Deep-learning super-resolution faces computation and memory demands that restrict deployment on resource-limited mobile devices.
Source: Introduction · p. 1Method
The method searches networks built from efficient residual dense blocks with an evolutionary algorithm that uses block credits to guide mutation.
Source: Introduction · p. 2, Efficient Residual Dense Blocks · p. 2, Evolutionary Algorithm · p. 4Results
The searched models outperform state-of-the-art methods with limited parameters and FLOPs; on Urban100 at ×2, medium and mini ESRN outperform FALSR by 0.44 PSNR and 0.51 PSNR, respectively.
Source: Introduction · p. 2, Comparison with State-of-the-art Methods · pp. 6–7Takeaways & Limitations
The approach yields fast, lightweight, accurate super-resolution networks by exploiting feature-scale variation and guiding evolution with block credits.
Source: Introduction · p. 2, Conclusion · p. 7
Abstract
from arXiv · showhide
Although remarkable progress has been made on single image super-resolution due to the revival of deep convolutional neural networks, deep learning methods are confronted with the challenges of computation and memory consumption in practice, especially for mobile devices. Focusing on this issue, we propose an efficient residual dense block search algorithm with multiple objectives to hunt for fast, lightweight and accurate networks for image super-resolution. Firstly, to accelerate super-resolution network, we exploit the variation of feature scale adequately with the proposed efficient residual dense blocks. In the proposed evolutionary algorithm, the locations of pooling and upsampling operator are searched automatically. Secondly, network architecture is evolved with the guidance of block credits to acquire accurate super-resolution network. The block credit reflects the effect of current block and is earned during model evaluation process. It guides the evolution by weighing the sampling probability of mutation to favor admirable blocks. Extensive experimental results demonstrate the effectiveness of the proposed searching method and the found efficient super-resolution models achieve better performance than the state-of-the-art methods with limited number of parameters and FLOPs.
Introduction
Single-image super-resolution has advanced with deep learning, but computation and memory demands constrain practical deployment, especially on mobile devices. The paper proposes efficient residual dense blocks and an evolutionary search guided by block credits to find fast, lightweight, accurate networks.
Introduction
- Deep super-resolution networks can be too costly for mobile devices; RDN, for example, has 22M parameters and requires 10,192G FLOPs.The paper identifies limited computing and storage resources as practical constraints.Source: Introduction · p. 1
- Hand-crafted super-resolution networks often repeat one block architecture, which may limit block flexibility and whole-network capacity.Prior work explored recursive, squeeze, and group-operation blocks, but the cited passage notes that uniform block designs may not be optimal.Source: Introduction · p. 1
- NAS had rarely been applied to lightweight super-resolution, while MoreMNAS searched cell and network architectures with basic operators and incurred substantial computation.The paper frames its search as an alternative that exploits feature-scale variation and regularized block choices.Source: Introduction · pp. 1–2
- The search combines three efficient residual dense block types with evolutionary architecture search and block credits that favor effective blocks during mutation.Pooling and upsampling locations are searched automatically; the blocks use local residual learning and global feature fusion to address pooling-related information loss.Source: Introduction · p. 2
- The authors report that their searched models outperform state-of-the-art methods while using limited parameters and FLOPs.The abstract attributes the approach to efficient blocks and block-credit-guided evolution.Source: Introduction · p. 2
Related Works
Prior super-resolution work pursued deeper networks, efficient architectures, and, more recently, neural architecture search. This paper positions its work among efforts to reduce model cost and automate architecture design.
Related Works
- Deep super-resolution progressed from SRCNN to deeper models using shortcuts and dense connections, including VDSR, RAAN, EDSR, and RDN.The passage identifies convolutional networks as a major direction in the field.Source: Related Works · p. 2
- Efficiency work delayed upsampling in FS-RCNN and ESPCN, used recursion to reduce parameters, and explored squeeze operations and group convolution.The cited examples include FS-RCNN (Dong, Loy, and Tang 2016), ESPCN (Shi et al. 2016), recursive methods (Tai, Yang, and Liu 2017; Tai et al. 2017), and lightweight models (Ahn, Kang, and Sohn 2018; Hui, Wang, and Gao 2018).Source: Related Works · p. 2
- Neural architecture search was also applied to super-resolution, with MoreMNAS cited as a recent example (Chu et al. 2019).The passage describes NAS as an additional direction alongside hand-designed efficient models.Source: Related Works · p. 2
Super-resolution Neural Architecture Search
The paper searches super-resolution backbones with evolution-based NAS and three lean residual dense block designs. It uses block credits to guide mutation and evaluates accuracy alongside computation and model size.
Super-resolution Neural Architecture Search
- The method uses evolution-based NAS to seek fast, lightweight, accurate super-resolution networks, with efficient residual dense blocks exploiting feature-scale variation.The architecture diagram depicts the proposed blocks within the search process.Source: Super-resolution Neural Architecture… · p. 2, Figure 1 · p. 3
Efficient Residual Dense Blocks
- Three lean residual dense blocks reduce cost through channel reduction, grouped convolution, or reduced feature scale.CRDB combines pooling with local residual connections and block-feature concatenation to retain information for reconstruction; recursion further expands its receptive field.Source: Efficient Residual Dense Blocks · p. 2
Efficiency Analysis
- Approximately 75% fewer FLOPs than RDB at R=1, while CRDB remains below RDB FLOPs at R=4.Pooling and recursion also expand the receptive field, and the search treats PSNR, FLOPs, and parameters as independent objectives.Source: Efficiency Analysis · p. 3
Efficient Residual Dense Block Search
- Because super-resolution search evaluates feature maps at the same or higher resolution, its model-evaluation procedure is especially computation-intensive.The authors motivate using characteristics of super-resolution to alleviate this search cost.Source: Efficient Residual Dense Block Search · p. 3
Search Space
- The search automatically varies backbone block count, block type, and block hyperparameters, including choices that determine low-resolution feature locations and convolution counts.Each block is encoded by its state, type, and hyperparameters.Source: Search Space · p. 3
Evolutionary Algorithm
- Block credit measures fitness gain from adding a block after preceding blocks and guides roulette-wheel mutation toward higher-credit architectures.Credits are approximated and updated during evolution rather than computed for every combination of preceding blocks.Source: Evolutionary Algorithm · p. 4
Performance Evaluation
- Evaluation uses PSNR for accuracy alongside parameter count and FLOPs; search can optimize a Pareto front or maximize PSNR under resource caps.The stated constraints place upper bounds on network parameters and FLOPs.Source: Performance Evaluation · pp. 4–5
Experiments
Experiments evaluate the proposed blocks and searched models using DIV2K and four benchmark datasets. ESRN and ESRN-V outperform prominent methods at similar parameter and Multi-Adds budgets, while ablations show that CRDB count and arrangement matter.
Datasets and Implementation
- Models are trained on DIV2K and evaluated on Set5, Set14, B100, and Urban100 using PSNR and SSIM on the Y channel.The evolution stage uses 32×32 LR patches, while retraining uses 64×64 LR patches.Source: Datasets and Implementation · p. 5
Comparison with State-of-the-art Methods
- ESRN and ESRN-V outperform prominent methods across benchmark datasets at similar parameter and Multi-Adds budgets.The comparisons cover ×2, ×3, and ×4 super-resolution scales.Source: Table 3 · p. 6, Comparison with State-of-the-art Methods · p. 6, Table 4 · p. 7
Ablation Study of Efficient Block
- More than five CRDBs cause PSNR to decrease rapidly, despite pooling and recursive operators expanding the receptive field.The reported ablation also finds that CRDB improves performance under parameter and Multi-Adds constraints.Source: Figure 3 · p. 5, Ablation Study of Efficient Block · p. 6
- First-four and cross arrangements of CRDBs perform better than placing the last four CRDBs together.The authors use these results to motivate searching block types and parameters for each block.Source: Table 1 · p. 5, Ablation Study of Efficient Block · p. 6
Comparison with State-of-the-art Methods
- On Set14’s Barbara image at ×3, ESRN reconstructions have clearer edges and fewer visible artifacts than the compared methods.The comparison is qualitative and is shown in Figure 5.Source: Figure 5 · p. 7, Comparison with State-of-the-art Methods · p. 7
- ESRN-F performs slightly better than FALSR-A on all four benchmark datasets with approximately half the Multi-Adds.Pooling decouples parameter count from Multi-Adds in the search for this medium-size model.Source: Comparison with State-of-the-art Methods · p. 7
Conclusion
The proposed search algorithm uses efficient residual dense blocks and an evolutionary algorithm to find fast, lightweight, accurate super-resolution networks. The found networks demonstrate advantages over state-of-the-art methods under specific parameter and FLOP requirements.
Conclusion
- The found networks demonstrate advantages over state-of-the-art super-resolution methods under specific parameter and FLOP requirements.Source: Conclusion · p. 7
- Three efficient block designs reduce network parameters and computation through channel number, convolution filter, and feature scale.Source: Conclusion · p. 7
- The evolutionary algorithm exploits feature-scale variation to accelerate super-resolution networks.Source: Conclusion · p. 7
- Block credit measures block effectiveness and guides evolution for searching acceleration.Source: Conclusion · p. 7