基于多尺度卷积神经网络的单幅图像去雾方法 下载: 2145次
Single Image Dehazing Method Based on Multi-Scale Convolution Neural Network
兰州交通大学电子与信息工程学院, 甘肃 兰州 730070
图 & 表
图 1. 大气散射物理模型
Fig. 1. Physical model of atmospheric scattering
下载图片 查看原文
图 2. MSDN模型图
Fig. 2. MSDN model diagram
下载图片 查看原文
图 3. 激活函数比较。(a) ReLU激活函数;(b) PReLU激活函数
Fig. 3. Comparison of activation functions. (a) ReLU activation function; (b) PReLU activation function
下载图片 查看原文
图 4. 本文算法步骤
Fig. 4. Algorithmic steps in this paper
下载图片 查看原文
图 5. 训练数据集。(a) ITS室内数据集;(b) OTS室外数据集
Fig. 5. Training data set. (a) Indoor data set ITS; (b) outdoor data set OTS
下载图片 查看原文
图 6. 合成有雾图像的实验结果。(a)有雾图像;(b)标准无雾图像;(c)文献[
7]方法;(d)文献[
11]方法;(e)文献[
12]方法;(f)文献[
13]方法;(g)文献[
14]方法;(h)本文方法
Fig. 6. Experimental results of synthesizing hazy images. (a) Hazy image; (b) standard haze-free image; (c) method in Ref. [7]; (d) method in Ref. [11]; (e) method in Ref. [12]; (f) method in Ref. [13]; (g) method in Ref. [14]; (h) proposed method
下载图片 查看原文
图 7. 真实室外有雾图像的实验结果。 (a)有雾图像;(b)文献[
7]方法;(c)文献[
11]方法;(d)文献[
12]方法;(e)文献[
13]方法;(f)文献[
14]方法;(g)本文方法
Fig. 7. Experimental results of real outdoor hazy images. (a) Hazy images; (b) method in Ref.[7]; (c) method in Ref.[11]; (d) method in Ref.[12]; (e) method in Ref.[13]; (f) method in Ref.[14]; (e) proposed method
下载图片 查看原文
表 1多尺度特征提取卷积核参数表
Table1. Parameter table of multi-scale feature extraction kernel
Type | Conv |
---|
Filter size | 3×3 | 5×5 | 7×7 | Filter number | 5 | 5 | 5 | Pad | 0 | 0 | 0 | Stride | 1 | 1 | 1 |
|
查看原文
表 2合成有雾图像的实验结果数据分析
Table2. Analysis of experimental data of synthetic hazy images
ImageNo. | Method in Ref.[7] | Method in Ref.[11] | Method in Ref.[12] | Method in Ref.[13] | Method in Ref.[14] | Proposed method | | | | | |
---|
PSNR /dB | SSIM /% | PSNR /dB | SSIM /% | PSNR /dB | | | | | SSIM /% | PSNR /dB | SSIM /% | PSNR /dB | SSIM /% | PSNR /dB | SSIM /% |
---|
1 | 23.5357 | 85.26 | 17.5424 | 59.51 | 20.1796 | 72.37 | 26.1247 | 85.33 | 22.9888 | 79.48 | 28.8218 | 86.41 | 2 | 19.3044 | 80.04 | 17.7552 | 72.36 | 21.6655 | 84.07 | 22.1977 | 88.70 | 20.7442 | 87.39 | 23.6027 | 91.66 | 3 | 17.8051 | 79.60 | 16.9521 | 72.87 | 20.3959 | 81.23 | 22.7414 | 89.35 | 19.0486 | 84.07 | 26.0691 | 89.73 | 4 | 20.1277 | 82.78 | 19.0534 | 79.38 | 21.5623 | 84.42 | 21.8444 | 86.42 | 19.4202 | 81.68 | 24.3402 | 91.60 | 5 | 20.2825 | 81.85 | 21.1444 | 82.85 | 17.8848 | 79.29 | 27.6081 | 93.54 | 25.4410 | 90.61 | 29.1285 | 94.78 |
|
查看原文
表 3室外有雾图像的实验结果数据分析
Table3. Analysis of experimental data of outdoor hazy images
ImageNo. | Method in Ref.[7] | Method in Ref.[11] | Method in Ref.[12] | Method in Ref.[13] | Method in Ref.[14] | Proposed method | | | | | |
---|
IE | AG | IE | AG | IE | | | | | AG | IE | AG | IE | AG | IE | AG |
---|
1 | 7.0555 | 14.64 | 7.0652 | 18.34 | 7.3984 | 18.52 | 7.2445 | 17.22 | 7.4048 | 20.26 | 7.6821 | 23.32 | 2 | 7.5155 | 14.68 | 7.3049 | 17.04 | 7.8192 | 18.83 | 7.4186 | 13.93 | 7.6656 | 17.02 | 7.9266 | 18.99 | 3 | 7.3427 | 8.48 | 7.4737 | 10.89 | 7.8771 | 11.78 | 7.7043 | 8.62 | 7.6120 | 9.21 | 7.8935 | 12.08 | 4 | 7.5688 | 9.18 | 7.4250 | 11.93 | 7.8515 | 13.23 | 7.7608 | 9.62 | 7.7786 | 10.57 | 7.9815 | 14.31 | 5 | 7.2538 | 14.91 | 7.7213 | 18.17 | 7.3410 | 17.93 | 7.1746 | 10.43 | 7.3776 | 13.62 | 7.8690 | 19.41 | 6 | 7.1667 | 8.92 | 7.8371 | 10.40 | 7.7168 | 10.18 | 7.0263 | 7.85 | 7.2862 | 8.76 | 7.8937 | 10.47 | 7 | 7.2625 | 10.94 | 7.1136 | 12.96 | 7.6834 | 13.71 | 7.3200 | 8.67 | 7.5189 | 9.59 | 7.7274 | 14.06 | 8 | 6.2721 | 5.88 | 7.1459 | 7.32 | 7.3561 | 7.55 | 6.8363 | 5.97 | 6.9818 | 6.39 | 7.4534 | 7.57 |
|
查看原文
表 4不同实验图像的算法运行时间
Table4. Running time of different algorithms for experimental imagess
Method | Experiment |
---|
Indoor | Outdoot |
---|
Method in Ref.[7] | 6.87 | 6.89 | Method in Ref.[11] | 3.41 | 3.65 | Method in Ref.[12] | 1.96 | 2.08 | Method in Ref.[13] | 1.21 | 1.26 | Method in Ref.[14] | 1.58 | 1.92 | Proposed method | 1.09 | 1.18 |
|
查看原文
陈永, 郭红光, 艾亚鹏. 基于多尺度卷积神经网络的单幅图像去雾方法[J]. 光学学报, 2019, 39(10): 1010001. Yong Chen, Hongguang Guo, Yapeng Ai. Single Image Dehazing Method Based on Multi-Scale Convolution Neural Network[J]. Acta Optica Sinica, 2019, 39(10): 1010001.