Acta Optica Sinica, Volume. 38, Issue 10, 1010002(2018)
Depth Map Super-Resolution Based on Two-Channel Convolutional Neural Network
Figures & Tables(10)
Fig. 1. Two-channel convolutional neural network for depth map super-resolution model
Fig. 4. Comparison of reconstruction results obtained by different methods. (a) Original image; (b) bicubic algorithm; (c) SRCNN algorithm; (d) proposed algorithm
Table 1. Parameter settings of the model
View tableTable 1. Parameter settings of the model
Name C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 DC1 DC1 Number 64 64 16 12 12 4 4 4 4 1 1 4 8 Size 3×3 1×1 1×1 3×3 1×1 7×7 5×5 3×3 1×1 5×5 1×1 15×15 15×15 Pad 1 0 0 1 0 3 2 1 0 2 0 7 7 Step 1 1 1 1 1 1 1 1 1 1 1 4/8 4/8
Table 2. Quantitative comparison of different algorithms on Art,Books and Moebius (RMSE)
View tableTable 2. Quantitative comparison of different algorithms on Art,Books and Moebius (RMSE)
Algorithm Art Books Moebius 4× 8× 4× 8× 4× 8× Bilinear 4.147 5.995 1.673 2.394 1.499 2.198 Ref.[23] 3.794 5.503 1.546 2.209 1.439 2.054 Ref.[13] 4.056 4.712 1.701 1.949 1.386 1.820 Ref.[15] 3.498 4.165 1.530 1.994 1.349 1.804 Ref.[12] 3.788 4.974 1.572 2.097 1.434 1.878 Ref.[18] 3.785 4.787 1.603 1.992 1.458 1.914 Ref.[22] - 5.798 - 2.728 - 2.422 Ref.[25] 2.525 3.957 1.098 1.646 0.979 1.459 Proposed 1.613 2.185 1.195 1.448 1.198 1.457
Table 3. Quantitative comparison of different algorithms on Dolls,Laundry and Reindeer (RMSE)
View tableTable 3. Quantitative comparison of different algorithms on Dolls,Laundry and Reindeer (RMSE)
Algorithm Dolls Laundry Reindeer 4× 8× 4× 8× 4× 8× Bicubic 1.305 1.855 2.408 3.452 2.809 3.986 Ref.[15] 1.301 1.745 2.132 2.770 2.407 2.987 Ref.[9] 1.977 - 2.969 - 3.178 - CLMF0[ 28 ]1.271 1.878 2.312 3.084 2.690 3.417 CLMF1[ 28 ]1.267 1.707 2.512 2.892 2.699 3.331 Ref.[18] 1.355 1.859 2.511 3.757 2.712 3.789 Ref.[21] 0.921 1.259 1.212 2.077 1.559 2.583 Ref.[25] 0.989 1.445 1.630 2.466 1.914 2.878 Proposed 0.842 1.462 1.285 1.547 1.237 1.600
Table 4. Quantitative comparison of different algorithms on Venus,Teddy and Cones (RMSE)
View tableTable 4. Quantitative comparison of different algorithms on Venus,Teddy and Cones (RMSE)
Algorithm Venus Teddy Cones 4× 8× 4× 8× 4× 8× Ref.[9] 2.597 - 4.030 - 5.740 - Ref.[10] 2.331 - 3.718 - 5.490 - Ref.[21] 0.819 1.169 1.822 2.370 2.974 4.516 Ref.[18] 1.742 - 2.595 - 3.498 - Ref.[22] 1.734 2.134 2.723 3.468 3.985 5.344 SRCNN[ 24 ]0.789 1.706 1.985 3.252 3.585 5.180 SRCNN2 0.718 1.593 1.891 3.136 3.439 5.171 Ref.[25] 1.191 1.786 2.026 3.015 3.078 4.865 Proposed 0.531 0.749 1.375 1.780 1.448 1.898
Table 5. Quantitative comparison of different algorithms by a scaling factor of 4
View tableTable 5. Quantitative comparison of different algorithms by a scaling factor of 4
Dataset Bicubic algorithm SRCNN algorithm[ 24 ]Proposed algorithm PSNR SSIM PSNR SSIM PSNR SSIM Art 29.81 0.8799 31.20 0.9260 31.82 0.9562 Books 30.22 0.9283 31.33 0.9425 31.88 0.9639 Dolls 31.67 0.9179 32.44 0.9300 32.61 0.9438 Laundry 32.64 0.9189 34.24 0.9330 34.37 0.9442 Moebius 30.92 0.9258 31.89 0.9403 32.52 0.9536 Reindeer 31.39 0.9290 33.30 0.9519 33.86 0.9647 Cones 28.61 0.8993 29.81 0.9190 30.57 0.9048 Teddy 27.61 0.9167 28.52 0.9343 29.70 0.9474 Venus 42.45 0.9827 45.52 0.9898 46.19 0.9923
Table 6. [in Chinese]
View tableTable 6. [in Chinese]
Algorithm Art Books Dolls Laundry Moebius Reindeer Cones Teddy Venus Bicubic 0.0398 0.0045 0.0034 0.0048 0.0035 0.0036 0.0042 0.0040 0.0036 SRCNN 0.2613 0.2580 0.2550 0.2459 0.2519 0.2456 0.2492 0.2479 0.2498 Proposed 0.2747 0.2008 0.1690 0.1624 0.2261 0.2107 0.1670 0.2324 0.1697
Tools
Get Citation
Copy Citation Text
Sumei Li, Guoqing Lei, Ru Fan. Depth Map Super-Resolution Based on Two-Channel Convolutional Neural Network[J]. Acta Optica Sinica, 2018, 38(10): 1010002
Paper Information
Category: Image Processing
Received: Feb. 6, 2018
Accepted: May. 16, 2018
Published Online: May. 9, 2019
The Author Email:
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20