Fig. 1. DSSR (top) and LapSRN (bottom) algorithms deal with different images. (a) Images processed with different scale factors; (b) building edges result
Fig. 2. Structure of DSSR
Fig. 3. Structure of dense block
Fig. 4. Structure of squeeze module
Fig. 5. Example of image data augmentation. (a) Original image; (b) horizontal flip image; (c) rotate 180° image; (d) vertical flip image
Fig. 6. Image decomposition map. (a) Decomposed image of Y channel; (b) decomposed image of Cb channel; (c) decomposed image of Cr channel
Fig. 7. Nonlinear output of different images. (a) Nonlinear output of nature image; (b) nonlinear output of text image; (c) nonlinear output of building image; (d) nonlinear output of person
Fig. 8. Nonlinear feature maps of feature extraction and image reconstruction. (a) Nonlinear mapping part feature map of the first convolution layer; (b) nonlinear mapping part feature map of the first concatenate layer
Fig. 9. Super-resolution results of “img_020” (bsd100) with scale factor ×2. (a) Original image; (b) result from Bicubic[28]; (c) result from A+[37]; (d) result from SRCNN[12]; (e) result from WSD-SR[38]; (f) result from VDSR[15]; (g) result from DRCN[39]; (h) result from LapSRN[40]; (i) result from DCSCN[19]; (j) DSSR
Fig. 10. Super-resolution results of “img_003” (Set5) with scale factor ×3. (a) Original image; (b) result from Bicubic; (c) result from A+; (d) result from SRCNN; (e) result from WSD-SR; (f) result from VDSR; (g) result from DRCN; (h) result from LapSRN; (i) result from DCSCN; (j) DSSR
Fig. 11. Super-resolution results of “img_012” (Set14) with scale factor×4. (a) Original image; (b) result from Bicubic; (c) result from A+; (d) result from SRCNN; (e) result from WSD-SR; (f) result from VDSR; (g) result from DRCN; (h) result from LapSRN; (i) result from DCSCN; (j) DSSR
Dataset | 3 DBs | 2 DBs(filters: 128) | 2DBs |
---|
Set5 | 33.65 | 33.76 | 33.87 | Set14 | 29.70 | 29.82 | 29.87 | B100 | 28.80 | 28.85 | 28.90 | Urban100 | 27.17 | 27.21 | 27.31 |
|
Table 1. Some parameters fine-tuned (number of dense blocks, number of filters) when the scale factor is 3. Bold fonts indicate the best performance
Parameter | SRCNN | VDSR | DRCN | LapSRN | DCSCN | DSSR |
---|
Number of CNN layers | 3 | 20 | 20 | 24 | 11 | 10 | Network input | LR with bicubic | LR with bicubic | LR with bicubic | LR | LR | LR | Residual learning | No | Yes | No | Yes | Yes | Yes | Loss function | L2 | L2 | L2 | Charbonnier | L2 | L1 | Activation function | ReLU | ReLU | ReLU | ReLU | ReLU | PReLU | Parameter number /k | 57 | 665 | 1775 | 812 | 87 | 72 |
|
Table 2. Technical implementation and parameter details of each super-resolution algorithm
Algorithm | Scale | Set5 | Set14 | B100 | Urban100 |
---|
PSNR /time | PSNR /time | PSNR /time | PSNR /time |
---|
BicubicA+WSD-SRSRCNNVDSRDRCNLapSRNDCSCNDSSR | 222222222 | 33.66/0.00 s36.54/0.58 s 37.21/0.34 s36.66/2.19 s37.53/0.13 s37.63/-37.52/0.058 s37.64/0.12 s37.73/0.010 s | 30.24/0.00 s32.26/0.86 s32.83/0.75 s32.42/4.32 s33.03/0.25 s33.06/-33.08/0.174 s33.05/0.21 s33.15/0.154 | 29.56/0.00 s31.21/0.59 s31.41/0.45 s31.36/2.51 s31.90/0.16 s31.85/-31.08/0.09 s31.91/0.18 s32.02/0.088 s | 26.88/0.00 s29.20/2.96 s30.29/2.96 s29.50/22.12 s30.76/0.98 s30.76/-30.41/0.752 s30.75/1.51 s30.96/0.780 s | BicubicA+WSD-SRSRCNNVDSRDRCNLapSRNDCSCNDSSR | 333333333 | 30.39/0.00 s32.58/0.32 s33.50/0.27 s32.75/2.23 s33.66/0.13 s33.85/-33.82/0.097 s33.75/0.51 s33.89/0.147 s | 27.55/0.00 s29.13/0.56 s29.72/0.34 s29.28/4.40 s29.77/0.26 s29.89/-29.87/0.13 s29.80/0.21 s29.90/0.139 s | 27.21/0.00 s28.29/0.33 s28.53/0.79 s28.41/2.58 s28.82/0.21 s28.81/-28.82/0.105 s28.80/0.14 s28.90/0.065 s | 24.46/0.00 s26.03/1.67 s26.95/1.41 s26.24/19.35 s27.14/1.08 s27.16/-27.07/0.57 s27.14/1.35 s27.17/0.878 s | BicubicA+WSD-SRSRCNNVDSRDRCNLapSRNDCSCNDSSR | 444444444 | 28.42/0.00 s30.28/0.24 s31.39/0.44 s30.48/2.19 s31.35/0.12 s31.56/-31.54/0.11 s31.40/0.22 s31.67/0.145 s | 26.00/0.00 s27.32/0.38 s27.98/0.48 s27.49/4.39 s28.01/0.25 s28.15/-28.19/0.20 s28.02/0.44 s28.21/0.12 s | 25.96/0.00 s26.82/0.26 s27.08/0.43 s26.90/2.51 s27.29/0.21 s27.24/-27.32/0.15 s27.31/0.36 s27.37/0.059 s | 23.14/0.00 s24.32/1.21 s25.16/1.36 s24.52/18.46 s25.18/1.06 s25.18/-25.21/0.56 s25.20/1.47 s25.24/0.952 s |
|
Table 3. Average PSNR and time of different scale factors on the four benchmark datasets Set5, Set14, B100 and Urban100 (Bold fonts indicate the best performance)