Fast, Accurate and Lightweight Super-Resolution with Neural Architecture Search

By Xiangxiang Chu, Bo Zhang, Hailong Ma, Ruijun Xu, Jixiang Li, Qingyuan Li. 2019

287 citations recorded by the source corpus.

Deep convolutional neural networks demonstrate impressive results in the super-resolution domain. A series of studies concentrate on improving peak signal noise ratio (PSNR) by using much deeper layers, which are not friendly to constrained resources. Pursuing a trade-off between the restoration capacity and the simplicity of models is still non-trivial. Recent contributions are struggling to manually maximize this balance, while our work achieves the same goal automatically with neural architecture search. Specifically, we handle super-resolution with a multi-objective approach. We also propose an elastic search tactic at both micro and macro level, based on a hybrid controller that profits from evolutionary computation and reinforcement learning. Quantitative experiments help us to draw a conclusion that our generated models dominate most of the state-of-the-art methods with respect to the individual FLOPS.

Subfields: Biomedical Engineering, Computer Vision and Pattern Recognition, Neural Architecture Search & AutoML

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Biomedical Engineering

1. Qi Tian
2. Hong Liu
3. James T. Kwok
4. Xu Jia
5. Chunhua Shen
6. Chen Gao
7. Ying Li
8. Yulun Zhang

In Computer Vision and Pattern Recognition

1. Luc Van Gool
2. Chunhua Shen
3. Qi Tian
4. Dacheng Tao
5. Ming-Hsuan Yang
6. Alan L. Yuille
7. Yi Yang
8. Xiaogang Wang

In Neural Architecture Search & AutoML

1. Bo Zhang
2. Xiangxiang Chu
3. Qi Tian
4. Bing Xue
5. Mengjie Zhang
6. Yanan Sun
7. Xin Chen
8. Gary G. Yen

Citations in the corpus

• Hierarchical Neural Architecture Search for Single Image Super-Resolution (2020, methodology): “To further show the effectiveness of the proposed method, we also conduct visual comparisons between HNAS and three SR methods, such as CARN [31], MoreMNAS [32] and FALSR [33].”
• Memory-Efficient Hierarchical Neural Architecture Search for Image Denoising (2019, methodology): “Two more relevant works are E-CAE [32] and FALSR [3].”
• Progressive Differentiable Architecture Search: Bridging the Depth Gap Between Search and Evaluation (2019, background): “Recently, the development of neural architecture search (NAS) has changed the convention of model design from manual to automatic, achieving remarkable success in various perceptual tasks [15, 3, 36] including image recognition [37].”
• Lightweight Image Super-Resolution with Adaptive Weighted Learning Network (2019, methodology): “Specifically, AWSRN-M shows better performance than SelfNet, MoreMNAS-A [13], FALSR-A [14] with similar parameter numbers and Muti-Adds.”
• Efficient Transformer for Single Image Super-Resolution (2021, methodology): “1) Objective Evaluation: In TABLE I, we compare our ESRT with 13 state-of-the-art lightweight SR models, including SRCNN [14], FSRCNN [15], VDSR [16], DRCN [3], LapSRN [43], DRRN [2], MemNet [4], EDSR-baseline [17], SRMDNF [44], CARN [7], FALSR [5], IMDN [11], and LatticeNet [23].”
• DARTS-: Robustly Stepping out of Performance Collapse Without Indicators (2020, methodology): “On the other hand, it also breaks the unfair advantage (Chu et al., 2020c) as the advantageous contribution from the residual block is factored out.”
• Bridging the Gap between Sample-based and One-shot Neural Architecture Search with BONAS (2020, background): “Results competitive with hand-crafted architectures have been obtained in many application areas, such as natural language processing [20, 30] and computer vision [25, 9, 3, 16, 4].”
• Multi-objective Neural Architecture Search via Predictive Network Performance Optimization (2019, unknown): “We also compare our method in open domains with NASNet search space (Zoph et al., 2018) and ResNet style search space (Li et al., 2019).”
• IR-NAS: Neural Architecture Search for Image Restoration (2019, background): “Compared with FALSR, our IR-NAS is 96 × fast in searching.”
• A Survey on Evolutionary Neural Architecture Search (2020, methodology): “Using an RNN to control the mutation operation can also be seen in other methods such as [129].”
• LatticeNet: Towards Lightweight Image Super-Resolution with Lattice Block (2020, methodology): “Besides, FALSR [6] applies neural architecture search to SR and performs excellently.”
• Efficient Residual Dense Block Search for Image Super-Resolution (2019, methodology): “Different from MoreMNAS (Chu et al. 2019), we add pooling and recursive operator in the search space, which decouples the FLOPs and parameter number.”
• NAS-DIP: Learning Deep Image Prior with Neural Architecture Search (2020, methodology): “NAS algorithms have been successfully applied to various tasks, including image classification [53, 61, 63, 64, 89, 90], semantic segmentation [14,51,58], object detection [24], image restoration [17,72], and image generation [26].”