基于多尺度融合和对抗训练的图像去雾算法 
下载： 1416次
Image Dehazing Algorithm Based on Multi-Scale Fusion and Adversarial Training
天津大学电气自动化与信息工程学院, 天津 300072
图 & 表
图 1. 生成器网络结构
Fig. 1. Architecture of the generative network
下载图片 查看原文
图 2. 多尺度特征提取模块示意图
Fig. 2. Illustration of the multi-scale feature extraction block
下载图片 查看原文
图 3. 残差密集连接模块示意图
Fig. 3. Illustration of the residual-and-densely-connected block
下载图片 查看原文
图 4. 鉴别器网络结构
Fig. 4. Architecture of the discriminative network
下载图片 查看原文
图 5. 合成有雾图像的去雾结果。(a)有雾图;(b) DCP;(c) FVR;(d) BCCR;(e) GRM;(f) CAP;(g) NLD;(h) DehazeNet;(i) MSCNN;(j) AOD-Net;(k)本算法;(l)真实无雾图
Fig. 5. Dehazing results of the synthetic hazy images. (a) Hazy image; (b) DCP; (c) FVR; (d) BCCR; (e) GRM; (f) CAP; (g) NLD; (h) DehazeNet; (i) MSCNN; (j) AOD-Net; (k) proposed method; (l) haze-free image
下载图片 查看原文
图 6. 真实场景有雾图的去雾结果。(a)有雾图;(b) DCP;(c) FVR;(d) BCCR;(e) GRM;(f) CAP;(g) NLD;(h) DehazeNet;(i) MSCNN;(j) AOD-Net;(k)本算法
Fig. 6. Dehazing results of the real-world hazy images. (a) Hazy image; (b) DCP; (c) FVR; (d) BCCR; (e) GRM; (f) CAP; (g) NLD; (h) DehazeNet; (i) MSCNN; (j) AOD-Net; (k) proposed method
下载图片 查看原文
表 1合成有雾图的全参考质量指标比较
Table1. Comparison of full-reference quality metrics tested on synthetic hazy images
| Quality metric | DCP | FVR | BCCR | GRM | CAP | NLD | DehazeNet | MSCNN | AOD-Net | Proposed |
|---|
| PSNR/dB | 16.62 | 15.72 | 16.88 | 18.86 | 19.05 | 17.29 | 21.14 | 17.57 | 19.06 | 25.08 | | SSIM | 0.8179 | 0.7483 | 0.7913 | 0.8553 | 0.8364 | 0.7489 | 0.8472 | 0.8102 | 0.8504 | 0.9468 |
|
查看原文
表 2合成有雾图的无参考质量指标比较
Table2. Comparison of no-reference quality metrics tested on synthetic hazy images
| Quality metric | DCP | FVR | BCCR | GRM | CAP | NLD | DehazeNet | MSCNN | AOD-Net | Proposed |
|---|
| SSEQ | 64.94 | 67.75 | 65.83 | 63.30 | 64.69 | 67.46 | 65.46 | 65.31 | 67.65 | 68.07 | | BLIINDS-II | 74.41 | 75.63 | 74.45 | 73.46 | 73.41 | 74.85 | 71.71 | 74.34 | 79.02 | 81.96 |
|
查看原文
表 3不同雾浓度的合成有雾图像的全参考质量指标比较
Table3. Comparison of full-reference quality metrics tested on synthetic hazy images with different haze concentration levels
| Hazedensity | Qualitymetric | DCP | FVR | BCCR | GRM | CAP | NLD | Dehaze-Net | MSCNN | AOD-Net | Ourproposed |
|---|
| Light hazeβ∈[0.6,0.9] | PSNR /dB | 16.10 | 17.18 | 16.91 | 18.64 | 20.88 | 17.52 | 24.24 | 19.72 | 22.40 | 26.11 | | SSIM | 0.8158 | 0.7682 | 0.7978 | 0.8528 | 0.8597 | 0.7558 | 0.9044 | 0.8489 | 0.8980 | 0.9579 | | Medium hazeβ∈[1.0,1.4] | PSNR /dB | 16.58 | 16.00 | 17.07 | 18.74 | 19.68 | 17.37 | 22.02 | 17.25 | 19.61 | 24.95 | | SSIM | 0.8210 | 0.7538 | 0.7942 | 0.8576 | 0.8450 | 0.7487 | 0.8870 | 0.8110 | 0.8616 | 0.9472 | | Heavy hazeβ∈[1.5,1.8] | PSNR /dB | 17.15 | 14.42 | 17.14 | 19.11 | 17.21 | 17.06 | 18.67 | 15.10 | 16.16 | 24.40 | | SSIM | 0.8259 | 0.7289 | 0.7906 | 0.8555 | 0.8120 | 0.7438 | 0.8454 | 0.7723 | 0.8064 | 0.9381 |
|
查看原文
表 4真实场景有雾图的无参考质量指标比较
Table4. Comparison of no-reference quality metrics tested on real-world hazy images
| Quality metric | DCP | FVR | BCCR | GRM | CAP | NLD | DehazeNet | MSCNN | AOD-Net | Proposed |
|---|
| SSEQ | 68.65 | 67.75 | 66.63 | 70.19 | 67.67 | 67.96 | 68.34 | 68.44 | 70.05 | 69.99 | | BLIINDS-II | 69.35 | 72.10 | 68.55 | 79.60 | 63.55 | 70.80 | 60.35 | 62.65 | 74.75 | 82.25 |
|
查看原文
刘宇航, 吴帅. 基于多尺度融合和对抗训练的图像去雾算法[J]. 激光与光电子学进展, 2020, 57(6): 061015. Yuhang Liu, Shuai Wu. Image Dehazing Algorithm Based on Multi-Scale Fusion and Adversarial Training[J]. Laser & Optoelectronics Progress, 2020, 57(6): 061015.
PDF全文
