基于显著性强度和梯度先验的多尺度图像盲去模糊 下载: 1002次
Multi-Scale Image Blind Deblurring Based on Salient Intensity and a priori Gradient
1 河海大学物联网工程学院, 江苏 常州 213022
2 中国工程物理研究院流体物理研究所, 四川 绵阳 621900
图 & 表
图 1. 模拟模糊图像(σ=3%)复原效果的定性比较。(a)模糊图像;(b)文献[
7]算法;(c)文献[
8]算法;(d)文献[
9]算法;(e)文献[
10]算法;(f)文献[
11]算法;(g)本文算法
Fig. 1. Qualitative comparison of restoration results for simulated blurred images (σ=3%). (a) Blurred image; (b) algorithm in Ref. [7]; (c) algorithm in Ref. [8]; (d) algorithm in Ref. [9]; (e) algorithm in Ref. [10]; (f) algorithm in Ref. [11]; (g) our algorithm
下载图片 查看原文
图 2. 不同算法的模糊核相似度比较
Fig. 2. Comparison of image blur kernel similarity under different algorithms
下载图片 查看原文
图 3. 真实图像复原效果的定性比较。 (a)模糊图像;(b)文献[
7]算法;(c)文献[
8]算法;(d)文献[
9]算法;(e)文献[
10]算法;(f)文献[
11]算法;(g)本文算法
Fig. 3. Qualitative comparison of restoration results for true images. (a) Blurred image; (b) algorithm in Ref. [7]; (c) algorithm in Ref. [8]; (d) algorithm in Ref. [9]; (e) algorithm in Ref. [10]; (f) algorithm in Ref. [11]; (g) our algorithm
下载图片 查看原文
表 1不同算法的PSNR和MSE统计结果
Table1. Statistic results of PSNR and MSE of different algorithms
Image | Algorithmin Ref. [7] | Algorithmin Ref. [8] | Algorithmin Ref. [9] | Algorithmin Ref. [10] | Algorithmin Ref. [11] | Our algorithm |
---|
PSNR /dB | MSE /% | PSNR /dB | MSE /% | PSNR /dB | MSE /% | PSNR /dB | MSE /% | PSNR /dB | MSE /% | PSNR /dB | MSE /% |
---|
Carton | 68.47 | 0.94 | 69.87 | 0.78 | 64.75 | 2.53 | 68.72 | 0.92 | 70.16 | 0.73 | 71.62 | 0.62 | House | 75.05 | 0.21 | 70.79 | 0.62 | 72.15 | 0.42 | 78.85 | 0.14 | 83.12 | 0.09 | 82.83 | 0.10 | Roman | 68.76 | 0.89 | 68.54 | 1.05 | 74.50 | 0.23 | 68.60 | 0.92 | 77.41 | 0.16 | 78.36 | 0.14 | Buddha | 74.42 | 0.25 | 71.59 | 0.67 | 73.38 | 0.33 | 74.34 | 0.26 | 75.13 | 0.23 | 76.85 | 0.21 | Building | 64.18 | 1.48 | 67.42 | 1.02 | 68.35 | 0.93 | 69.74 | 0.81 | 70.22 | 0.71 | 72.47 | 0.58 |
|
查看原文
表 2使用自适应迭代前后运行时间统计结果
Table2. Statistical results before and after using adaptive iteration
Romanimage | Layer 7image | Layer 6image | Layer 5image | Layer 4image | Layer 3image | Layer 2image | Layer 1image | Totaltime |
---|
Image resolution | 10×7 | 19×14 | 38×27 | 75×53 | 150×105 | 300×210 | 600×420 | | Kernel resolution | 7×7 | 11×11 | 15×15 | 21×21 | 29×29 | 39×39 | 51×51 | | Normal iteration /s | 18 | 36 | 90 | 221 | 429 | 864 | 1944 | 3602 | Adaptive iteration /s | 18 | 36 | 82 | 160 | 306 | 558 | 1026 | 2196 |
|
查看原文
表 36种算法的运行时间(Roman图)
Table3. Run time of 6 algorithms (Roman image)
Romanimage | Algorithmin Ref. [7] | Algorithmin Ref. [8] | Algorithmin Ref. [9] | Algorithmin Ref. [10] | Algorithmin Ref. [11] | Our algorithm(normalization) | Our algorithm(adaptiveiteration) |
---|
Time /h | 0.10 | 0.12 | 0.43 | 0.92 | 0.95 | 1.00 | 0.61 |
|
查看原文
陈晨, 许金鑫, 危才华, 李庆武. 基于显著性强度和梯度先验的多尺度图像盲去模糊[J]. 激光与光电子学进展, 2020, 57(4): 041505. Chen Chen, Jinxin Xu, Caihua Wei, Qingwu Li. Multi-Scale Image Blind Deblurring Based on Salient Intensity and a priori Gradient[J]. Laser & Optoelectronics Progress, 2020, 57(4): 041505.