Fast, Accurate, and, Lightweight Super-Resolution with Cascading Residual Network

Namhyuk AhnByungkon KangKyung-ah Sohn

article2018ECCV1,446 citations

Proposes a cascading residual network for single-image super-resolution that matches state-of-the-art accuracy while dramatically cutting parameters and computational cost to enable practical real-world deployment.

Listen

Single-image super-resolution reconstructs high-resolution images from low-resolution inputs, playing a vital role in consumer mobile applications, video streaming services, and surveillance systems. While deep learning methods have drastically improved image restoration quality, state-of-the-art architectures demand substantial computational operations and memory. This computational weight creates severe operational bottlenecks, including high latency and excessive battery consumption on edge devices.

The main objective of the article is to develop and evaluate high-performing, lightweight deep learning architectures that significantly reduce computational operations and parameter sizes while preserving image reconstruction accuracy. To accomplish this, the authors introduce the Cascading Residual Network (CARN) and its mobile-optimized variant (CARN-M).

The approach designs neural network blocks that connect intermediate layers across both local and global cascading pathways, upsampling the image only at the final stage to avoid heavy intermediate calculations. To create the mobile variant, the authors incorporate group convolutions into an efficient residual block and share parameters recursively. The models were trained on the standard DIV2K dataset using the L1 loss function and evaluated against established benchmarks across standard test sets, including Set5, Set14, B100, and Urban100, at multiple scaling factors.

The findings demonstrate that CARN outperforms existing models of comparable size (under five million parameters) across all standard benchmarks while maintaining a modest computational burden of 90.9 billion multiply-accumulate operations for 720p resolution at 4x scaling. The streamlined CARN-M variant reduces the parameter count by approximately 74% (from 1,592K to 412K) and cuts computational operations by roughly 64% (from 90.9G to 32.5G) relative to CARN, while incurring only a slight drop of 0.29 dB in peak signal-to-noise ratio. The ablation analysis revealed that combining local and global cascading is essential: using local cascading alone degraded performance due to optimization bottlenecks, whereas global shortcuts restored effective information and gradient flow. Furthermore, multi-scale learning allowed a single trained model to handle multiple upscaling factors simultaneously.

These results show that high visual quality does not require massive computing infrastructure. By decoupling image enhancement from excessive processing demands, organizations can deploy real-time super-resolution on resource-constrained mobile hardware and significantly lower bandwidth and storage expenses in media streaming via on-the-fly decompression. Unlike prior recursive approaches that maintained accuracy only by increasing network depth and latency, the proposed cascading framework achieves high fidelity with high execution efficiency.

For practical implementation, engineering teams should adopt CARN when image fidelity is paramount within a moderate compute budget, and deploy CARN-M when targeting battery-sensitive mobile platforms or ultra-low-latency streaming pipelines. Future development should focus on extending this cascading methodology directly to real-time video streaming architectures to validate end-to-end compression and decompression workflows.

The findings carry high confidence across standard static image benchmarks, though decision-makers should note that evaluations were conducted primarily on standard benchmark image sets. Performance across continuous, real-time video streams under varying network conditions remains subject to future empirical validation.

Ajou University deep learning image super-resolution super-resolution reconstruction lightweight super-resolution cascading residual network single image super-resolution

arXiv: 1803.08664 nmhkahn/CARN-pytorch

Paper: Deep Residual Learning for Image Recognition, Kaiming He et al. (2016). It establishes deep residual learning and skip connections, which form the architectural foundation for the cascading residual network.
Paper: Image Super-Resolution Using Deep Convolutional Networks, Chao Dong et al. (2014). It pioneers end-to-end deep learning frameworks for single-image super-resolution that subsequent lightweight networks aim to optimize.
Paper: Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network, Wenzhe Shi et al. (2016). It introduces sub-pixel convolution for efficient upsampling in the low-resolution domain, a key technique for reducing computational cost in super-resolution.
Paper: Accurate Image Super-Resolution Using Very Deep Convolutional Networks, Jiwon Kim et al. (2016). It integrates deep residual learning into super-resolution networks, motivating deeper yet parameter-efficient SR architectures.
Paper: Accelerating the Super-Resolution Convolutional Neural Network, Chao Dong et al. (2016). It redesigns convolutional super-resolution for acceleration by performing feature extraction directly in the low-resolution space.
Paper: Image Super-Resolution via Deep Recursive Residual Network, Ying Tai et al. (2017). It demonstrates recursive residual learning to boost super-resolution depth while strictly constraining parameter growth.
Paper: Enhanced Deep Residual Networks for Single Image Super-Resolution, Bee Lim et al. (2017). It optimizes residual architectures specifically for super-resolution by removing unnecessary modules like batch normalization.
Paper: Deep Laplacian Pyramid Networks for Fast and Accurate Super-Resolution, Wei-Sheng Lai et al. (2017). It proposes a progressive, cascaded pyramid framework for fast and accurate single-image super-resolution.

Paper: Image Super-Resolution Using Very Deep Residual Channel Attention Networks, Yulun Zhang et al. (2018). It advances deep residual super-resolution architectures by incorporating residual-in-residual structures and channel attention mechanisms.
Paper: Residual Dense Network for Image Super-Resolution, Yulun Zhang et al. (2018). It extends hierarchical feature reuse in super-resolution by combining dense connections with local and global residual learning.
Paper: Second-Order Attention Network for Single Image Super-Resolution, Tao Dai et al. (2019). It builds on deep residual super-resolution by integrating second-order channel attention and non-local operations.
Paper: Deep Learning for Image Super-Resolution: A Survey, Zhihao Wang et al. (2019). It provides a comprehensive survey synthesizing deep learning architectures, lightweight strategies, and benchmarks for image super-resolution.
Paper: SwinIR: Image Restoration Using Swin Transformer, Jingyun Liang et al. (2021). It transitions the field from convolutional residual designs to shifted-window transformer backbones for classical and lightweight super-resolution.
Paper: Multi-Stage Progressive Image Restoration, Syed Waqas Zamir et al. (2021). It extends multi-stage progressive restoration strategies across diverse degradation tasks using cross-stage feature propagation.

Cover for Fast, Accurate, and, Lightweight Super-Resolution with Cascading Residual Network

Abstract

In recent years, deep learning methods have been successfully applied to single-image super-resolution tasks. Despite their great performances, deep learning methods cannot be easily applied to real-world applications due to the requirement of heavy computation. In this paper, we address this issue by proposing an accurate and lightweight deep network for image super-resolution. In detail, we design an architecture that implements a cascading mechanism upon a residual network. We also present variant models of the proposed cascading residual network to further improve efficiency. Our extensive experiments show that even with much fewer parameters and operations, our models achieve performance comparable to that of state-of-the-art methods.

Topics

deep learning image super-resolution super-resolution reconstruction network architecture network components

Keywords

lightweight super-resolution cascading residual network single image super-resolution deep residual learning CARN model local cascading global cascading efficient residual block image restoration depthwise convolution

1 Introduction
2 Related Work
2.1 Deep Learning Based Image Super-Resolution
2.2 Efficient Neural Network
3 Proposed Method
3.1 Cascading Residual Network
3.2 Efficient Cascading Residual Network
3.3 Comparison to Recent Models
4 Experimental Results
4.1 Datasets
4.2 Implementation and Training Details
4.3 Comparison with State-of-the-art Methods
4.4 Model Analysis
Cascading Modules.
Efficiency Trade-off.
5 Conclusion
References

Knowls

Knowl 1 — Cascading Residual Network (CARN) Architecture
model/method
The Cascading Residual Network (CARN) is a deep convolutional neural network designed for fast, accurate, and lightweight single-image super-resolution (SISR). Instead of upsampling the input low-resolution (LR) image at the beginning of the network, CARN extracts and processes features entirely at the LR resolution and upsamples to high resolution (HR) at the end of the network using sub-pixel convolution (pixel shuffle).

The CARN architecture consists of three main parts:
1. Initial Feature Extraction: A single convolutional layer extracts shallow feature representations from the input LR image $X$ .
2. Cascading Backbone (Global and Local Cascading): A sequence of $B=3$ cascading blocks with global cascading connections connecting the output of the initial convolution and each preceding cascading block into subsequent cascading blocks. Inside each cascading block, $U=3$ residual blocks are arranged with local cascading connections. At both global and local cascading junctions, intermediate multi-level representations are concatenated channel-wise and compressed using a $1\times 1$ convolutional layer.
3. Multi-Scale Reconstruction and Upsampling: The aggregated multi-level feature map is fed into scale-specific upsampling modules composed of convolutions and pixel shuffle operations to generate the HR image for scaling factors $\times 2$ , $\times 3$ , or $\times 4$ .
By integrating cascading mechanisms at both global (layer-wise) and local (block-wise) levels, CARN facilitates multi-level representation learning and creates multiple shortcut paths that assist gradient flow during optimization.
Knowl 2 — Mathematical Formulation of Multi-Level Cascading
equation
Let $f(\cdot; W)$ denote a convolution operation parameterized by weights $W$ , and $\tau(\cdot)$ denote the ReLU activation function.

A standard residual block $R_i$ with parameter set $W_R^i = \{W_R^{i,1}, W_R^{i,2}\}$ operating on an input feature map $H^{i-1}$ is defined as: $R_i(H^{i-1}; W_R^i) = \tau\left(f\left(\tau\left(f\left(H^{i-1}; W_R^{i,1}\right)\right); W_R^{i,2}\right) + H^{i-1}\right)$

The $i$ -th local cascading block $B_{\text{local}}^i(H^{i-1}; W_l^i) \equiv B^{i,U}$ containing $U$ residual units is defined recursively for $u = 1, \dots, U$ by: $B^{i,0} = H^{i-1}$ $B^{i,u} = f\left(\left[B^{i,0}, B^{i,1}, \dots, B^{i,u-1}, R^u\left(B^{i,u-1}; W_R^u\right)\right]; W_c^{i,u}\right)$ where $[\cdot]$ represents channel-wise feature concatenation and $W_c^{i,u}$ denotes the parameters of the $1\times 1$ convolution compressing the concatenated features.

The global cascading process over $B$ cascading blocks computes intermediate representations $H^b$ for $b = 1, \dots, B$ from an input image $X$ as: $H^0 = f(X; W_c)$ $H^b = f\left(\left[H^0, H^1, \dots, H^{b-1}, B_{\text{local}}^b\left(H^{b-1}; W_B^b\right)\right]; W_g^b\right)$ where $W_c$ is the parameter set of the initial convolution layer and $W_g^b$ represents the parameters of the $1\times 1$ convolution merging all preceding global representations. In CARN, $U = 3$ and $B = 3$ .
Knowl 3 — Efficient Residual Block (residual-E) and Computation Reduction Ratio
equation
The efficient residual block (residual-E) replaces standard convolutions in a residual block with group convolutions and a pointwise ( $1\times 1$ ) convolution to make computational cost tunable.

A residual-E block consists of two $3\times 3$ group convolutions with group size $G$ , followed by one $1\times 1$ pointwise convolution. Let $K$ denote the kernel size ( $K=3$ ), $C_{\text{in}}$ and $C_{\text{out}}$ denote the number of input and output channels (kept constant at 64), and $F \times F$ denote the spatial resolution of the input and output feature maps.

The computational cost (in multiply-accumulate operations) of a standard residual block is: $\text{Cost}_{\text{standard}} = 2 \times \left(K \cdot K \cdot C_{\text{in}} \cdot C_{\text{out}} \cdot F \cdot F\right)$

The computational cost of a residual-E block is: $\text{Cost}_{\text{residual-E}} = 2 \times \left(K \cdot K \cdot C_{\text{in}} \cdot \frac{C_{\text{out}}}{G} \cdot F \cdot F\right) + C_{\text{in}} \cdot C_{\text{out}} \cdot F \cdot F$

The theoretical computation reduction ratio of the residual-E block relative to the standard residual block is: $\frac{\text{Cost}_{\text{residual-E}}}{\text{Cost}_{\text{standard}}} = \frac{2 \times \left(K \cdot K \cdot C_{\text{in}} \cdot \frac{C_{\text{out}}}{G} \cdot F \cdot F\right) + C_{\text{in}} \cdot C_{\text{out}} \cdot F \cdot F}{2 \times \left(K \cdot K \cdot C_{\text{in}} \cdot C_{\text{out}} \cdot F \cdot F\right)} = \frac{1}{G} + \frac{1}{2K^2}$

For $K=3$ , the reduction ratio is $\frac{1}{G} + \frac{1}{18}$ , reducing convolution computation by a factor between $1.8\times$ (for $G=2$ ) and $14\times$ (for $G=64$ ).
Knowl 4 — CARN-Mobile (CARN-M) Architecture
model/method
CARN-Mobile (CARN-M) is an ultra-lightweight variant of the Cascading Residual Network designed to optimize both parameter footprint and floating-point operations (Mult-Adds) for mobile and real-time super-resolution applications.

CARN-M alters the base CARN design via two modifications:
1. Efficient Residual Blocks (residual-E): Standard residual blocks inside all cascading blocks are replaced with residual-E blocks, setting the group convolution group size to $G=4$ .
2. Recursive Parameter Sharing: The parameters across all cascading blocks are shared, converting the global cascading backbone into a recursive cascading architecture. Parameter sharing reduces the parameter count of the cascading blocks by up to three times.
Through these modifications, CARN-M reduces parameter count from 1,592K (CARN) down to 412K (a nearly $4\times$ reduction) and cuts Mult-Adds by approximately $2.5\times$ to $2.8\times$ across $\times 2$ , $\times 3$ , and $\times 4$ scales, with only a modest 0.29 dB average PSNR penalty relative to CARN.
Knowl 5 — Uniform Weight Initialization for Cascading Networks with 1x1 Convolutions
model/method
Standard weight initialization routines (such as Xavier/Glorot or He initialization) assign excessively large initial weight values to narrow $1\times 1$ convolutional layers present in cascading shortcuts, which causes numerical instability and training failure in cascading architectures.

To ensure stable convergence, all weights and biases $\theta$ across CARN and CARN-M are initialized from a uniform distribution: $\theta \sim \mathcal{U}(-k, k), \quad \text{with } k = \frac{1}{\sqrt{c_{\text{in}}}}$ where $c_{\text{in}}$ is the number of input channels to the respective convolutional layer. This initialization prevents exploding activation and gradient values during backpropagation through cascading connections.

Knowl 6 — Quantitative SISR Benchmark Comparison

data/table

CARN and CARN-M were benchmarked against state-of-the-art deep learning SISR models on Set5, Set14, B100, and Urban100 datasets across scaling factors $\times 2$ , $\times 3$ , and $\times 4$ . Multiply-accumulate operations (Mult-Adds) are calculated assuming a $720\text{p}$ ( $1280 \times 720$ ) output image.

Scale	Model	Params	Mult-Adds	Set5 PSNR/SSIM	Set14 PSNR/SSIM	B100 PSNR/SSIM
$\times 2$	SRCNN	57K	52.7G	36.66 / 0.9542	32.42 / 0.9063	31.36 / 0.8879
$\times 2$	FSRCNN	12K	6.0G	37.00 / 0.9558	32.63 / 0.9088	31.53 / 0.8920
$\times 2$	VDSR	665K	612.6G	37.53 / 0.9587	33.03 / 0.9124	31.90 / 0.8960
$\times 2$	DRCN	1,774K	17,974.3G	37.63 / 0.9588	33.04 / 0.9118	31.85 / 0.8942
$\times 2$	DRRN	297K	6,796.9G	37.74 / 0.9591	33.23 / 0.9136	32.05 / 0.8973
$\times 2$	MemNet	677K	2,662.4G	37.78 / 0.9597	33.28 / 0.9142	32.08 / 0.8978
$\times 2$	SelNet	974K	225.7G	37.89 / 0.9598	33.61 / 0.9160	32.08 / 0.8984
$\times 2$	CARN (ours)	1,592K	222.8G	37.76 / 0.9590	33.52 / 0.9166	32.09 / 0.8978
$\times 2$	CARN-M (ours)	412K	91.2G	37.53 / 0.9583	33.26 / 0.9141	31.92 / 0.8960
$\times 3$	SRCNN	57K	52.7G	32.75 / 0.9090	29.28 / 0.8209	28.41 / 0.7863
$\times 3$	FSRCNN	12K	5.0G	33.16 / 0.9140	29.43 / 0.8242	28.53 / 0.7910
$\times 3$	VDSR	665K	612.6G	33.66 / 0.9213	29.77 / 0.8314	28.82 / 0.7976
$\times 3$	DRRN	297K	6,796.9G	34.03 / 0.9244	29.96 / 0.8349	28.95 / 0.8004
$\times 3$	SelNet	1,159K	120.0G	34.27 / 0.9257	30.30 / 0.8399	28.97 / 0.8025
$\times 3$	CARN (ours)	1,592K	118.8G	34.29 / 0.9255	30.29 / 0.8407	29.06 / 0.8034
$\times 3$	CARN-M (ours)	412K	46.1G	33.99 / 0.9236	30.08 / 0.8367	28.91 / 0.8000
$\times 4$	SRCNN	57K	52.7G	30.48 / 0.8628	27.49 / 0.7503	26.90 / 0.7101
$\times 4$	FSRCNN	12K	4.6G	30.71 / 0.8657	27.59 / 0.7535	26.98 / 0.7150
$\times 4$	VDSR	665K	612.6G	31.35 / 0.8838	28.01 / 0.7674	27.29 / 0.7251
$\times 4$	DRRN	297K	6,796.9G	31.68 / 0.8888	28.21 / 0.7720	27.38 / 0.7284
$\times 4$	SelNet	1,417K	83.1G	32.00 / 0.8931	28.49 / 0.7783	27.44 / 0.7325
$\times 4$	SRDenseNet	2,015K	389.9G	32.02 / 0.8934	28.50 / 0.7782	27.53 / 0.7337
$\times 4$	CARN (ours)	1,592K	90.9G	32.13 / 0.8937	28.60 / 0.7806	27.58 / 0.7349
$\times 4$	CARN-M (ours)	412K	32.5G	31.92 / 0.8903	28.42 / 0.7762	27.44 / 0.7304

On Urban100 ( $\times 4$ ), CARN achieves $26.07\text{ dB} / 0.7837$ SSIM with $90.9\text{G}$ Mult-Adds, outperforming SRDenseNet ( $26.05\text{ dB} / 0.7819$ at $389.9\text{G}$ Mult-Adds) and DRRN ( $25.44\text{ dB} / 0.7638$ at $6,796.9\text{G}$ Mult-Adds). CARN-M achieves $25.62\text{ dB} / 0.7694$ SSIM on Urban100 ( $\times 4$ ) with only $32.5\text{G}$ Mult-Adds and $412\text{K}$ parameters.

Knowl 7 — Ablation of Local and Global Cascading Modules

data/table

An ablation study on Set14 ( $\times 4$ scale) isolates the contribution of local and global cascading mechanisms relative to a plain ResNet baseline.

Architecture	Baseline (ResNet)	CARN-NL	CARN-NG	CARN
Local Cascading	-	-	✓	✓
Global Cascading	-	✓	-	✓
# Parameters	1,444K	1,481K	1,555K	1,592K
PSNR (dB)	28.43	28.45	28.42	28.52

Key findings:

Local cascading alone degrades performance: CARN-NG (local cascading only) drops to 28.42 dB PSNR, performing worse than the plain baseline (28.43 dB). Multiplicative $1\times 1$ convolutions placed along local shortcut paths hinder information propagation and gradient flow during training.
Global cascading aids multi-level propagation: CARN-NL (global cascading only) achieves 28.45 dB PSNR by effectively transferring multi-level frequency signals from shallow to deep layers.
Dual cascading synergy: Combining local and global cascading (CARN) attains the highest performance (28.52 dB PSNR). Global cascading connections provide direct bypasses that resolve the optimization bottleneck caused by local $1\times 1$ convolutions, allowing the network to fully leverage multi-level local feature representations.

Knowl 8 — Group Size and Recursion Trade-offs in CARN-M
empirical result
Analyzing group sizes $G \in \{2, 4, 8, 16, 32, 64\}$ in efficient residual blocks (residual-E) across non-recursive and recursive configurations on Set14 ( $\times 4$ ) demonstrates the following:
1. Group Size vs. Performance: Increasing group size $G$ reduces Mult-Adds and parameters monotonically. For example, $G=64$ achieves a $5\times$ reduction in operations and parameters compared to $G=1$ , but incurs noticeable PSNR degradation.
2. Effect of Recursive Sharing: Applying recursive weight sharing across cascading blocks reduces the parameter count by up to $3\times$ without changing the total Mult-Adds. For a fixed parameter budget, recursive models consistently achieve higher PSNR than non-recursive models with larger group sizes.
3. Optimal Configuration (G4R): Setting group size $G=4$ with recursive cascading blocks (designated G4R) delivers the optimal trade-off. Relative to full CARN, CARN-M (G4R) reduces parameters by $3.8\times$ (from 1,592K to 412K) and operations by $2.8\times$ (from 90.9G to 32.5G Mult-Adds) with a minor PSNR loss of 0.29 dB.
Knowl 9 — Training Methodology and Multi-Scale Training Setup
experimental setup
CARN and CARN-M are trained under the following unified experimental protocol:
- Dataset: DIV2K dataset (800 training high-resolution images). Training inputs are $64 \times 64$ RGB patches randomly cropped from LR images, augmented with random horizontal flips and $90^\circ$ rotations.
- Loss Function: $L_1$ loss between reconstructed and ground truth HR patches. Compared to $L_2$ loss, $L_1$ provides superior convergence stability and higher reconstruction quality in residual SISR architectures.
- Optimizer and Hyperparameters: ADAM optimizer with $\beta_1 = 0.9$ , $\beta_2 = 0.999$ , and $\epsilon = 10^{-8}$ . Minibatch size is 64. The model is trained for $6 \times 10^5$ iterations with an initial learning rate of $10^{-4}$ , halved every $4 \times 10^5$ iterations.
- Multi-Scale Learning: Mini-batches cycle through scaling factors $\times 2$ , $\times 3$ , and $\times 4$ . A single trained CARN(-M) network shares feature extraction layers across all scales and activates the scale-specific sub-pixel upsampling head corresponding to the target scale, producing a unified multi-scale model file of only 1.6 MB for CARN-M.

Coverage note — Visual qualitative comparisons (Fig. 6) and background comparisons to SRDenseNet/MemNet were omitted as standalone knowls as their core insights are subsumed by the benchmark data and model architecture knowls.

References

1.Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2017)
2.Arbelaez, P., Maire, M., Fowlkes, C., Malik, J.: Contour detection and hierarchical image segmentation. IEEE transactions on pattern analysis and machine intelligence 33(5), 898–916 (2011)
3.Bevilacqua, M., Roumy, A., Guillemot, C., Alberi-Morel, M.: Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In: Proceedings of the British Machine Vision Conference (BMVC) (2012)
4.Choi, J.S., Kim, M.: A deep convolutional neural network with selection units for super-resolution. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2017)
5.Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2009)
6.Dong, C., Loy, C.C., He, K., Tang, X.: Learning a deep convolutional network for image super-resolution. In: Proceedings of the European Conference on Computer Vision (ECCV) (2014)
7.Dong, C., Loy, C.C., Tang, X.: Accelerating the super-resolution convolutional neural network. In: Proceedings of the European Conference on Computer Vision (ECCV) (2016)
8.Fan, Y., Shi, H., Yu, J., Liu, D., Han, W., Yu, H., Wang, Z., Wang, X., Huang, T.S.: Balanced two-stage residual networks for image super-resolution. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2017)
9.Girshick, R.: Fast r-cnn. In: Proceedings of the International Conference on Computer Vision (ICCV) (2015)
10.Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: Proceedings of the International Conference on Artificial Intelligence and Statistics (2010)
11.Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. Proceedings of the International Conference on Learning Representations (ICLR) (2016)
12.He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: Proceedings of the International Conference on Computer Vision (ICCV) (2015)
13.He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2016)
14.He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2016)
15.He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: Proceedings of the European Conference on Computer Vision (ECCV) (2016)
16.Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)
17.Huang, G., Liu, Z., van der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2017)
18.Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2015)
19.Iandola, F.N., Han, S., Moskewicz, M.W., Ashraf, K., Dally, W.J., Keutzer, K.: Squeezenet: Alexnet-level accuracy with 50x fewer parameters and¡ 0.5 mb model size. arXiv preprint arXiv:1602.07360 (2016)
20.Kim, J., Kwon Lee, J., Mu Lee, K.: Accurate image super-resolution using very deep convolutional networks. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2016)
21.Kim, J., Kwon Lee, J., Mu Lee, K.: Deeply-recursive convolutional network for image super-resolution. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2016)
22.Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. Proceedings of the International Conference on Learning Representations (ICLR) (2015)
23.Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Proceedings of the Conference on Neural Information Processing Systems (NIPS) (2012)
24.Lai, W.S., Huang, J.B., Ahuja, N., Yang, M.H.: Deep laplacian pyramid networks for fast and accurate super-resolution. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2017)
25.Lee, J., Nam, J.: Multi-level and multi-scale feature aggregation using pretrained convolutional neural networks for music auto-tagging. IEEE Signal Processing Letters 24(8), 1208–1212 (2017)
26.Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2017)
27.Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.Y., Berg, A.C.: Ssd: Single shot multibox detector. In: Proceedings of the European Conference on Computer Vision (ECCV) (2016)
28.Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2015)
29.Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: Proceedings of the International Conference on Computer Vision (ICCV) (2001)
30.Noh, H., Hong, S., Han, B.: Learning deconvolution network for semantic segmentation. In: Proceedings of the International Conference on Computer Vision (ICCV) (2015)
31.Ren, H., El-Khamy, M., Lee, J.: Image super resolution based on fusing multiple convolution neural networks. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2017)
32.Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (2015)
33.Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2016)
34.Sifre, L., Mallat, S.: Rigid-motion scattering for image classification. Ph.D. thesis, Citeseer (2014)
35.Tai, Y., Yang, J., Liu, X.: Image super-resolution via deep recursive residual network. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2017)
36.Tai, Y., Yang, J., Liu, X., Xu, C.: Memnet: A persistent memory network for image restoration. In: Proceedings of the International Conference on Computer Vision (ICCV) (2017)
37.Tong, T., Li, G., Liu, X., Gao, Q.: Image super-resolution using dense skip connections. In: Proceedings of the International Conference on Computer Vision (ICCV) (2017)
38.Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600–612 (2004)
39.Yang, J., Wright, J., Huang, T.S., Ma, Y.: Image super-resolution via sparse representation. IEEE transactions on image processing 19(11), 2861–2873 (2010)
40.Zhang, R., Isola, P., Efros, A.A.: Colorful image colorization. In: Proceedings of the European Conference on Computer Vision (ECCV) (2016)

Citation

MLA

Ahn, N., et al. “Fast, Accurate, and Lightweight Super-Resolution with Cascading Residual Network”. arXiv, 2018, http://arxiv.org/abs/1803.08664v5.

APA

Ahn, N., Kang, B., & Sohn, K.-A. (2018). Fast, Accurate, and Lightweight Super-Resolution with Cascading Residual Network. arXiv. http://arxiv.org/abs/1803.08664v5

Chicago

Ahn, N., B. Kang, and K.-A. Sohn. 2018. “Fast, Accurate, and Lightweight Super-Resolution with Cascading Residual Network”. arXiv. http://arxiv.org/abs/1803.08664v5.

Harvard

Ahn, N., Kang, B. and Sohn, K.-A. (2018) “Fast, Accurate, and Lightweight Super-Resolution with Cascading Residual Network”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1803.08664v5.

Vancouver

1. Ahn N, Kang B, Sohn K-A (2018) Fast, Accurate, and Lightweight Super-Resolution with Cascading Residual Network. arXiv

BibTeX

@article{ahn2018fast,
  title = {Fast, Accurate, and Lightweight Super-Resolution with Cascading Residual Network},
  author = {Ahn, Namhyuk and Kang, Byungkon and Sohn, Kyung-Ah},
  year = {2018},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1803.08664v5},
  eprint = {1803.08664}
}

Metadata:arXiv

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

Fast, Accurate, and, Lightweight Super-Resolution with Cascading Residual Network

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Cascading Residual Network (CARN) Architecture

Knowl 2 — Mathematical Formulation of Multi-Level Cascading

Knowl 3 — Efficient Residual Block (residual-E) and Computation Reduction Ratio

Knowl 4 — CARN-Mobile (CARN-M) Architecture

Knowl 5 — Uniform Weight Initialization for Cascading Networks with 1x1 Convolutions