Acta Optica Sinica, Volume. 37, Issue 12, 1210002(2017)
Depth Map Super-Resolution Reconstruction Based on Convolutional Neural Networks
Figures & Tables(10)
Fig. 3. Comparison of reconstruction results obtained from different network layers
Fig. 4. Comparison of reconstruction results obtained by different methods. (a)-(c) Original depth maps and corresponding partially enlarged maps; (d)-(f) depth maps of bicubic interpolation algorithm and corresponding partially enlarged maps; (g)-(i) depth maps of SRCNN algorithm and corresponding partially enlarged maps; (j)-(l) depth maps of proposed algorithm and corresponding partially enlarged maps
Table 1. Parameters setting of each convolution layer
View tableTable 1. Parameters setting of each convolution layer
Name Number Size Pad Conv1 56 3×3 1 Conv2 12 1×1 0 Conv3 12 3×3 1 Conv4 12 3×3 1 Conv5 12 3×3 1 Conv6 12 3×3 1 Conv7 56 1×1 0
Table 2. Comparison of different network layers
View tableTable 2. Comparison of different network layers
Number of layers Time /s PSNR /dB 7 92.702 31.5321 10 113.076 33.4809 12 127.040 31.8638
Table 3. Quantitative comparison of different algorithms on the datasets
View tableTable 3. Quantitative comparison of different algorithms on the datasets
Depth map Bicubic interpolation algorithm SRCNN algorithm Proposed algorithm RMSE PSNR RMSE PSNR RMSE PSNR Art 1.5313 31.3049 1.3814 32.9155 1.3464 33.4809 Books 1.1334 31.2227 1.1324 31.7996 1.0385 33.7587 Moebius 1.1412 31.9929 1.1334 32.6431 1.0513 34.6130 Dolls 1.1052 32.6389 1.1043 33.1548 1.1017 34.7382 Laundry 1.1572 33.6228 1.0848 34.8337 1.0744 36.1265 Reindeer 1.1758 32.6560 1.0806 34.2127 1.0373 35.5582 Cones 1.4905 29.6001 1.4425 30.4998 1.3820 32.1335 Teddy 1.4659 28.8076 1.3950 29.6820 1.3430 31.3920 Venus 0.5610 43.8180 0.5076 45.7149 0.5014 48.3748
Table 4. Quantitative comparison 1 of different algorithms on the dataset
View tableTable 4. Quantitative comparison 1 of different algorithms on the dataset
Algorithm Art Books Moebius Bicubic 4.147 1.673 1.449 Ref. [18] 3.794 1.546 1.439 Ref. [5] 3.498 1.530 1.349 Ref. [9] 3.788 1.572 1.434 Ref. [11] 3.785 1.603 1.458 Proposed 1.692 1.248 1.257
Table 5. Quantitative comparison 2 of different algorithms on the dataset
View tableTable 5. Quantitative comparison 2 of different algorithms on the dataset
Algorithm Dolls Laundry Reindeer Bicubic 1.305 2.408 2.809 Ref. [5] 1.301 2.132 2.407 Ref. [19] 1.977 2.969 3.178 Ref. [11] 1.355 2.511 2.712 Ref. [8] 1.350 2.255 2.431 Ref. [22] 0.989 1.630 1.914 Proposed 1.274 1.311 1.275
Table 6. Quantitative comparison 3 of different algorithms on the dataset
View tableTable 6. Quantitative comparison 3 of different algorithms on the dataset
Algorithm Venus Teddy Cones Ref. [5] 1.8 4.89 5.64 Ref. [12] 3.55 4.92 6.34 Ref. [11] 2.52 3.3 4.45 Ref. [14] 0.82 1.82 2.97 Ref. [22] 1.19 2.03 3.08 Ref. [13] 1.14 1.80 2.13 Proposed 0.61 1.52 1.57
Tools
Get Citation
Copy Citation Text
Sumei Li, Guoqing Lei, Ru Fan. Depth Map Super-Resolution Reconstruction Based on Convolutional Neural Networks[J]. Acta Optica Sinica, 2017, 37(12): 1210002
Paper Information
Category: Image Processing
Received: Jun. 20, 2017
Accepted: --
Published Online: Sep. 6, 2018
The Author Email: Lei Guoqing (lgq20051118@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20