基于卷积神经网络的单幅图像超分辨 下载: 1324次
Single Image Super-Resolution Based on Convolutional Neural Network
图 & 表
图 1. SRCNN算法框架
Fig. 1. SRCNN algorithm framework
下载图片 查看原文
图 2. 本文算法框架图
Fig. 2. Proposed algorithm framework
下载图片 查看原文
图 3. 函数示意图。 (a) ReLU;(b) e-ReLU
Fig. 3. Function schematic. (a) ReLU; (b) e-ReLU
下载图片 查看原文
图 4. 训练过程中本文算法训练误差值随着迭代次数增加的变化图
Fig. 4. Graph of train loss in the proposed method with the increase of iterations in the training process
下载图片 查看原文
图 5. Set 5 中的baby_GT重建结果比较。(a)原图;(b) BI/33.91 dB;(c) ScSR/34.29 dB; (d) SRCNN/34.83 dB;(e) SRCNN-Ex/34.91 dB;(f)本文方法/35.04 dB
Fig. 5. Comparison of the reconstruction of the baby_GT in Set 5. (a) Original image; (b) BI/33.91 dB; (c) ScSR/34.29 dB; (d) SRCNN/34.83 dB; (e) SRCNN-Ex/34.91dB; (f) proposed method/35.04 dB
下载图片 查看原文
图 6. Set 5 中的butterfly_GT重建结果比较。(a)原图;(b) BI/24.04 dB;(c) ScSR/25.58 dB; (d) SRCNN/25.00 dB;(e) SRCNN-Ex/25.58 dB;(f)本文方法/27.91 dB
Fig. 6. Comparison of the reconstruction of the butterfly_GT in Set 5. (a) Original image; (b) BI/24.04 dB; (c) ScSR/25.58 dB; (d) SRCNN/25.00 dB; (e) SRCNN-Ex/25.58 dB; (f) proposed method/27.91 dB
下载图片 查看原文
图 7. Set 14 中的lenna重建结果比较。(a)原图;(b) BI/31.68 dB;(c) ScSR/32.64 dB; (d) SRCNN/32.53 dB;(e) SRCNN-Ex/32.78 dB;(f)本文方法/33.57 dB
Fig. 7. Comparison of the reconstruction of the lenna in Set 14. (a) Original image; (b) BI/31.68 dB; (c) ScSR/32.64 dB; (d) SRCNN/32.53 dB; (e) SRCNN-Ex/32.78 dB; (f) proposed method/33.57 dB
下载图片 查看原文
图 8. Set 14中的pepper重建结果比较。(a)原图;(b) BI /32.38 dB;(c) ScSR/33.32 dB; (d) SRCNN/32.08 dB;(e) SRCNN-Ex/33.30 dB;(f)本文方法/34.57 dB
Fig. 8. Comparison of the reconstruction of the pepper in Set 14. (a) Original image; (b) BI/32.38 dB; (c) ScSR/33.32 dB; (d) SRCNN/32.08 dB; (e) SRCNN-Ex/33.30 dB; (f) proposed method/34.57 dB
下载图片 查看原文
图 9. 随着迭代次数的增加,本文算法在Set 5测试集上平均PSNR值变化图
Fig. 9. Change graph of the average PSNR value for proposed algorithm in the Set 5 test set, with the number of iterations
下载图片 查看原文
表 1各个层的参数设置
Table1. Parameter settings for each layer
Name | Size | Number | Stride | Padding |
---|
Conv1 | 5×5 | 64 | 1 | 0 | Conv2 | 3×3 | 32 | 1 | 0 | Deconv | 9×9 | 1 | 3 | 4 |
|
查看原文
表 2在Set 5测试集上的PSNR和SSIM值
Table2. PSNR and SSIM values on Set 5 test set
Image | BI | ScSR | SRCNN | SRCNN-Ex | Proposed method |
---|
PSNR /dB | SSIM | PSNR /dB | SSIM | PSNR /dB | SSIM | PSNR /dB | SSIM | PSNR /dB | SSIM |
---|
Baby | 33.91 | 0.90 | 34.29 | 0.92 | 34.83 | 0.92 | 34.91 | 0.92 | 35.04 | 0.92 | Bird | 32.57 | 0.93 | 34.11 | 0.92 | 33.77 | 0.94 | 34.03 | 0.94 | 35.46 | 0.95 | Butterfly | 24.04 | 0.82 | 25.58 | 0.82 | 25.00 | 0.83 | 25.58 | 0.84 | 27.91 | 0.91 | Head | 32.88 | 0.80 | 33.17 | 0.80 | 33.42 | 0.82 | 33.42 | 0.82 | 33.67 | 0.83 | Women | 28.56 | 0.89 | 29.94 | 0.91 | 29.60 | 0.91 | 29.91 | 0.91 | 31.22 | 0.93 | Average | 30.39 | 0.87 | 31.42 | 0.87 | 31.32 | 0.88 | 31.57 | 0.89 | 32.66 | 0.91 |
|
查看原文
表 3在Set 14测试集上的PSNR和SSIM值
Table3. PSNR and SSIM values on Set 14 test set
Image | BI | ScSR | SRCNN | SRCNN-Ex | Proposed method |
---|
PSNR /dB | SSIM | PSNR /dB | SSIM | PSNR /dB | SSIM | PSNR /dB | SSIM | PSNR /dB | SSIM |
---|
Baboon | 23.21 | 0.54 | 23.50 | 0.59 | 23.52 | 0.60 | 23.54 | 0.60 | 23.62 | 0.61 | Barbara | 26.25 | 0.75 | 26.39 | 0.75 | 26.76 | 0.78 | 26.84 | 0.78 | 26.57 | 0.78 | Bridge | 24.40 | 0.65 | 24.80 | 0.70 | 24.89 | 0.70 | 24.95 | 0.70 | 25.14 | 0.71 | Coastguard | 26.55 | 0.61 | 27.00 | 0.65 | 27.00 | 0.66 | 27.08 | 0.66 | 27.12 | 0.66 | Comic | 23.12 | 0.70 | 23.90 | 0.76 | 23.77 | 0.75 | 23.87 | 0.75 | 24.53 | 0.79 | Face | 32.82 | 0.80 | 33.10 | 0.81 | 33.38 | 0.82 | 33.40 | 0.82 | 33.71 | 0.83 | Flowers | 27.23 | 0.80 | 28.25 | 0.83 | 28.06 | 0.83 | 28.27 | 0.83 | 29.22 | 0.85 | Foreman | 31.16 | 0.91 | 32.04 | 0.91 | 32.09 | 0.91 | 32.01 | 0.91 | 33.65 | 0.94 | Lenna | 31.68 | 0.86 | 32.64 | 0.87 | 32.53 | 0.87 | 32.78 | 0.88 | 33.57 | 0.88 | Man | 27.01 | 0.75 | 27.76 | 0.78 | 27.56 | 0.78 | 27.72 | 0.78 | 28.33 | 0.80 | Monarch | 29.43 | 0.92 | 30.71 | 0.93 | 30.40 | 0.93 | 30.87 | 0.93 | 32.78 | 0.95 | Pepper | 32.38 | 0.87 | 33.32 | 0.87 | 32.08 | 0.88 | 33.30 | 0.88 | 34.57 | 0.89 | Ppt3 | 23.71 | 0.87 | 24.98 | 0.87 | 24.34 | 0.88 | 25.02 | 0.89 | 26.24 | 0.92 | Zebra | 26.63 | 0.80 | 27.95 | 0.82 | 27.74 | 0.84 | 28.37 | 0.84 | 29.11 | 0.85 |
|
查看原文
表 4训练时间对比
Table4. Comparison of training times
Method | 1000 times iteration | 105 times iteration | 2×105 times iteration | 8×108 times iteration |
---|
SRCNN | 477 | | | 381600000 | SRCNN-Ex | 1392 | | | 1113600000 | Proposed method | 141 | 14100 | 28200 | |
|
查看原文
史紫腾, 王知人, 王瑞, 任福全. 基于卷积神经网络的单幅图像超分辨[J]. 激光与光电子学进展, 2018, 55(12): 121001. Ziteng Shi, Zhiren Wang, Rui Wang, Fuquan Ren. Single Image Super-Resolution Based on Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2018, 55(12): 121001.