基于深度学习的单幅图像去雾算法 
下载： 1481次
Single-Image Defogging Algorithm Based on Deep Learning
咸阳师范学院数学与信息科学学院, 陕西 咸阳 712000
图 & 表
图 1. 大气散射模型示意图
Fig. 1. Schematic of atmospheric scattering model
下载图片 查看原文
图 2. 所提网络结构
Fig. 2. Proposed network structure
下载图片 查看原文
图 3. 所提算法流程图
Fig. 3. Flow chart of proposed algorithm
下载图片 查看原文
图 4. 不同算法对雾天图像Teddy的去雾结果。(a)雾图;(b)清晰图像;(c) DCP算法;(d) BCCR算法;(e) SVDSR算法;(f) CAP算法;(g) MSCNN算法;(h)所提算法
Fig. 4. Defogging results of foggy image Teddy by different algorithms. (a) Foggy image; (b) original clear image; (c) DCP algorithm; (d) BCCR algorithm; (e) SVDSR algorithm; (f) CAP algorithm; (g) MSCNN algorithm; (h) proposed algorithm
下载图片 查看原文
图 5. 不同算法对雾天图像Dolls的去雾结果。(a)雾图;(b)清晰图像; (c) DCP算法;(d) BCCR算法;(e) SVDSR算法;(f) CAP算法;(g) MSCNN算法;(h)所提算法
Fig. 5. Defogging results of foggy image Dolls by different algorithms. (a) Foggy image; (b) original clear image; (c) DCP algorithm; (d) BCCR algorithm; (e) SVDSR algorithm; (f) CAP algorithm; (g) MSCNN algorithm; (h) proposed algorithm
下载图片 查看原文
图 6. 不同算法对雾天图像Cloth的去雾结果。(a)雾图;(b)清晰图像; (c) DCP算法;(d) BCCR算法;(e) SVDSR算法;(f) CAP算法;(g) MSCNN算法;(h)所提算法
Fig. 6. Defogging results of foggy image Cloth by different algorithms. (a) Foggy image; (b) original clear image; (c) DCP algorithm; (d) BCCR algorithm; (e) SVDSR algorithm; (f) CAP algorithm; (g) MSCNN algorithm; (h) proposed algorithm
下载图片 查看原文
图 7. 自然雾天图像House的去雾结果对比。(a)雾图;(b) DCP算法;(c) BCCR算法;(d) SVDSR算法;(e) CAP算法;(f) MSCNN算法;(g)所提算法
Fig. 7. Comparison of defogging results of natural foggy image House. (a) Foggy image; (b) DCP algorithm; (c) BCCR algorithm; (d) SVDSR algorithm; (e) CAP algorithm; (f) MSCNN algorithm; (g) proposed algorithm
下载图片 查看原文
图 8. 自然雾天图像Pumpkin的去雾结果对比。(a)雾图;(b) DCP算法; (c) BCCR算法;(d) SVDSR算法;(e) CAP算法;(f) MSCNN算法;(g)所提算法
Fig. 8. Comparison of defogging results of natural foggy image Pumpkin. (a) Foggy image; (b) DCP algorithm; (c) BCCR algorithm; (d) SVDSR algorithm; (e) CAP algorithm; (f) MSCNN algorithm; (g) proposed algorithm
下载图片 查看原文
图 9. 自然雾天图像Girls的去雾结果对比。(a)雾图;(b) DCP算法;(c) BCCR算法;(d) SVDSR算法;(e) CAP算法;(f) MSCNN算法;(g)所提算法
Fig. 9. Comparison of defogging results of natural foggy image Girls. (a) Foggy image; (b) DCP algorithm; (c) BCCR algorithm; (d) SVDSR algorithm; (e) CAP algorithm; (f) MSCNN algorithm; (g) proposed algorithm
下载图片 查看原文
图 10. 不同算法的对比结果。(a)平均梯度;(b)信息熵
Fig. 10. Comparison results of different algorithms. (a) Average gradient; (b) information entropy
下载图片 查看原文
表 1多尺度卷积参数
Table1. Multi-scale convolution parameters
| Filter size | Pad | Stride |
|---|
| 1×1×16 | 0 | 1 | | 3×3×165×5×167×7×16 | 123 | 111 |
|
查看原文
表 2图像Teddy采用不同算法去雾后评价指标结果
Table2. Evaluation indicators of defogging results of image Teddy by different algorithms
| Indicator | DCP | BCCR | SVDSR | CAP | MSCNN | Proposed |
|---|
| RMSE ↓UQI ↑Cross entropy ↑Tone reduction↑Average gradient ↑Entropy ↑PSNR /dB ↑ | 0.02730.61460.57630.756311.004417.026515.8246 | 0.01330.56781.13160.24328.806913.351312.5962 | 0.06290.62450.36430.75019.831814.628015.2510 | 0.02590.61791.63660.79087.149916.799519.8702 | 0.02580.60351.20880.69667.979916.385520.6022 | 0.02460.63271.25290.844111.195617.883123.4425 | | SSIM ↑ | 0.7782 | 0.6097 | 0.7572 | 0.8769 | 0.8797 | 0.9524 |
|
查看原文
表 3图像Dolls采用不同算法去雾后评价指标结果
Table3. Evaluation indicators of defogging results of image Dolls by different algorithms
| Indicator | DCP | BCCR | SVDSR | CAP | MSCNN | Proposed |
|---|
| RMSE ↓UQI ↑Cross entropy ↑Tone reduction↑Average gradient ↑Entropy ↑PSNR /dB↑ | 0.03200.59470.23000.93046.274614.980111.4845 | 0.02270.64402.54090.34556.956013.321810.6521 | 0.07560.67240.25360.76947.469613.741719.4985 | 0.03130.61591.37980.52363.949414.551824.6558 | 0.02970.59550.68060.55774.367114.313222.3259 | 0.02990.67812.58200.98437.556216.790824.7741 | | SSIM ↑ | 0.8412 | 0.6339 | 0.8601 | 0.8769 | 0.8583 | 0.9245 |
|
查看原文
表 4图像Cloth采用不同算法去雾后评价指标结果
Table4. Evaluation indicators of defogging results of image Cloth by different algorithms
| Indicator | DCP | BCCR | SVDSR | CAP | MSCNN | Proposed |
|---|
| RMSE ↓UQI ↑Cross entropy ↑Tone reduction↑Average gradient ↑Entropy ↑PSNR /dB↑ | 0.03750.83781.01920.745916.562913.314924.2385 | 0.02870.92201.43420.624322.515215.109816.2698 | 0.09660.88381.02650.63237.326912.659815.2502 | 0.02310.52310.22600.80455.614615.540523.4958 | 0.02410.68950.48520.65565.231214.231921.2102 | 0.02250.98670.04210.965022.768016.699527.3441 | | SSIM ↑ | 0.8567 | 0.7357 | 0.7279 | 0.9462 | 0.8975 | 0.9690 |
|
查看原文
赵建堂. 基于深度学习的单幅图像去雾算法[J]. 激光与光电子学进展, 2019, 56(11): 111005. Jiantang Zhao. Single-Image Defogging Algorithm Based on Deep Learning[J]. Laser & Optoelectronics Progress, 2019, 56(11): 111005.
PDF全文
