1 Introduction
2 Related works
3 Original DARTS for SR tasks
3.1 Preliminary of differentiable architecture search
3.2 The phenomenon of performance drop caused by intractable skip connections
4 Improved DARTS for super-resolution tasks
4.1 Remove redundant operations
4.2 Adding identity mapping on convolution operations
4.3 The architecture of the proposed network
5 Experiments and results
5.1 PSNR/SSIM for measuring reconstruction quality
5.2 Architecture search
5.3 Architecture evaluation
5.3.1 Evaluation on DIV2K dataset
5.3.2 Benchmark results
Dataset | Bicublic | A+ | SRCNN | VDSR | SRResNet | Original DARTS | Ours |
---|---|---|---|---|---|---|---|
Set5 | 28.42/0.8104 | 30.28/0.8603 | 30.48/0.8628 | 31.35/0.8838 | 32.05/0.8910 | 32.21/0.8949 | 32.21/0.8958 |
Set14 | 26.00/0.7027 | 27.32/0.7491 | 27.49/0.7503 | 28.01/0.7674 | 28.53/0.7804 | 28.66/0.7823 | 28.61/0.7832 |
B100 | 25.96/0.6675 | 26.82/0.7087 | 26.90/0.7101 | 27.29/0.7251 | 27.57/0.7354 | 27.51/0.7323 | 27.61/0.7384 |
Urban100 | 23.14/0.6577 | 24.32/0.7183 | 24.52/0.7221 | 25.18/0.7524 | 26.07/0.7839 | 25.96/0.7687 | 26.26/0.7932 |