基于生成对抗网络的多模态图像融合 下载: 2228次
Multimodal Image Fusion Based on Generative Adversarial Networks
1 中北大学大数据学院, 山西 太原 030051
2 酒泉卫星发射中心, 甘肃 酒泉 735000
图 & 表
图 1. 残差块结构图
Fig. 1. Structure of residual block
下载图片 查看原文
图 2. 方法框架图
Fig. 2. Framework of method
下载图片 查看原文
图 3. 生成模型的网络结构
Fig. 3. Network structure of generative model
下载图片 查看原文
图 4. 判别模型的网络结构
Fig. 4. Network structure of discriminative model
下载图片 查看原文
图 5. 标签图像预选图。(a)红外长波;(b)红外短波;(c)可见光;(d) LP;(e) DWT;(f) NSCT;(g) NSST
Fig. 5. Pre-selection maps of label images. (a) Longwave infrared; (b) shortwave infrared; (c) visible light; (d) LP; (e) DWT; (f) NSCT; (g) NSST
下载图片 查看原文
图 6. 学习率对生成器损失的影响
Fig. 6. Effect of learning rate on generator loss
下载图片 查看原文
图 7. 学习率对判别器损失的影响
Fig. 7. Effect of learning rate on discriminator loss
下载图片 查看原文
图 8. λ取值对图像质量的影响。(a) λ=0 ;(b) λ=0.01;(c) λ=0.1;(d) λ=1
Fig. 8. Effect of different λ on image quality. (a) λ=0; (b) λ=0.01; (c) λ=0.1; (d) λ=1
下载图片 查看原文
图 9. λ取值对生成器损失的影响
Fig. 9. Effect of different λ on generator loss
下载图片 查看原文
图 10. λ取值对融合图像客观评价指标的影响。(a)第1组融合图像;(b)第2组融合图像;(c)第3组融合图像
Fig. 10. Effect of λ on objective evaluation index of fused image. (a) The first set of fused images; (b) the second set of fused images; (c) the third set of fused images
下载图片 查看原文
图 11. 图像融合结果。(a)红外长波;(b)红外短波;(c) 可见光;(d) DTCWT_SR;(e) NSST_NSCT;(f) CNN;(g) CSR;(h)本文方法
Fig. 11. Image fusion results. (a) Longwave infrared; (b) shortwave infrared; (c) visible light; (d) DTCWT_SR; (e) NSST_NSCT; (f) CNN; (g) CSR; (h) proposed method
下载图片 查看原文
表 1生成器参数
Table1. Parameters of generator
Layer | Filter size /step | Output size |
---|
Conv1 | 3×3 /1 | 128×128×64 | Res(7 units) | 3×3 /1 | 128×128×64 | | 3×3 /1 | 128×128×64 | Conv9 | 3×3 /1 | 128×128×256 | Conv10 | 3×3 /1 | 128×128×1 |
|
查看原文
表 2判别器参数
Table2. Parameters of discriminator
Layer | Filter size/step | Output size |
---|
Conv1 | 3×3 /1 | 128×128×64 | Conv2 | 3×3 /2 | 64×64×128 | Conv3 | 3×3 /2 | 32×32×256 | Conv4 | 3×3 /2 | 16×16×512 | Conv5 | 3×3 /1 | 16×16×256 | Conv6 | 1×1 /1 | 16×16×128 | Res | 1×1 /1 | 16×16×64 | | 3×3 /1 | 16×16×64 | | 3×3 /1 | 16×16×128 | Fc | - | 1 |
|
查看原文
表 3标签图像选取依据表
Table3. Label image selection table
Fusion method | SD | AG | Con | CC | IE | MI | VIFF |
---|
LP | 49.019 | 4.211 | 40.812 | 0.424 | 7.223 | 5.608 | 0.466 | DWT | 44.772 | 3.826 | 36.650 | 0.342 | 7.175 | 5.415 | 0.442 | NSCT | 41.123 | 4.061 | 34.816 | 0.437 | 6.953 | 5.222 | 0.469 | NSST | 40.904 | 4.149 | 34.902 | 0.441 | 6.932 | 5.201 | 0.467 |
|
查看原文
表 4融合结果评价指标比较
Table4. Comparison of evaluation index of fusion results
Image | Fusion method | SD | AG | Con | CC | IE | MI | VIFF |
---|
No. 1 | DTCWT_SR | 35.471 | 3.150 | 23.916 | 0.409 | 6.968 | 4.869 | 0.505 | NSCT_NSST | 19.621 | 3.163 | 11.911 | 0.406 | 6.147 | 2.853 | 0.522 | CNN | 34.928 | 2.885 | 23.254 | 0.408 | 6.961 | 4.610 | 0.363 | CSR | 12.081 | 1.953 | 7.570 | 0.418 | 5.534 | 2.663 | 0.359 | Proposed method | 38.011 | 8.129 | 27.793 | 0.431 | 6.977 | 2.552 | 0.301 | No. 2 | DTCWT_SR | 26.274 | 7.245 | 25.927 | 0.143 | 5.984 | 1.492 | 0.367 | NSCT_NSST | 23.268 | 7.270 | 14.815 | 0.306 | 6.026 | 1.542 | 0.376 | CNN | 29.696 | 4.985 | 22.754 | 0.016 | 2.373 | 1.312 | 0.211 | CSR | 25.271 | 5.543 | 15.989 | 0.322 | 6.032 | 2.540 | 0.322 | Proposed method | 25.527 | 4.374 | 13.731 | 0.381 | 6.057 | 2.567 | 0.461 | No. 3 | DTCWT_SR | 21.854 | 5.187 | 15.038 | 0.427 | 6.475 | 1.256 | 0.415 | NSCT_NSST | 22.692 | 5.370 | 15.7166 | 0.471 | 6.806 | 1.542 | 0.419 | CNN | 38.590 | 4.961 | 25.609 | 0.441 | 6.910 | 2.799 | 0.392 | CSR | 23.054 | 3.868 | 15.726 | 0.499 | 6.450 | 2.391 | 0.382 | Proposed method | 41.089 | 4.310 | 30.929 | 0.593 | 6.938 | 3.093 | 0.420 | No. 4 | DTCWT_SR | 47.492 | 3.545 | 23.176 | 0.422 | 6.315 | 2.773 | 0.298 | NSCT_NSST | 35.373 | 3.605 | 15.136 | 0.445 | 6.831 | 2.531 | 0.317 | CNN | 54.597 | 3.163 | 29.778 | 0.436 | 6.283 | 2.862 | 0.351 | CSR | 38.587 | 2.217 | 17.656 | 0.452 | 6.796 | 3.331 | 0.375 | Proposed method | 32.400 | 8.452 | 22.767 | 0.457 | 6.938 | 2.055 | 0.384 | No. 5 | DTCWT_SR | 40.006 | 5.058 | 32.554 | 0.025 | 7.308 | 2.976 | 0.407 | NSCT_NSST | 24.886 | 5.147 | 18.059 | 0.489 | 6.615 | 1.812 | 0.428 | CNN | 40.559 | 3.951 | 34.331 | 0.271 | 7.316 | 2.133 | 0.346 | CSR | 26.239 | 2.526 | 20.707 | 0.515 | 6.642 | 3.007 | 0.277 | Proposed method | 40.677 | 5.974 | 33.890 | 0.569 | 7.321 | 2.106 | 0.473 | No. 6 | DTCWT_SR | 54.458 | 3.038 | 44.020 | 0.042 | 7.744 | 2.986 | 0.524 | NSCT_NSST | 25.990 | 3.047 | 18.728 | 0.455 | 6.686 | 2.254 | 0.552 | CNN | 45.549 | 2.538 | 35.269 | 0.193 | 7.333 | 1.839 | 0.208 | CSR | 28.803 | 1.523 | 21.125 | 0.429 | 6.691 | 2.858 | 0.160 | Proposed method | 46.425 | 5.643 | 36.462 | 0.476 | 7.453 | 2.993 | 0.648 | No. 7 | DTCWT_SR | 54.848 | 3.555 | 36.984 | 0.044 | 7.420 | 3.485 | 0.484 | NSCT_NSST | 27.527 | 3.550 | 21.490 | 0.460 | 6.813 | 1.807 | 0.504 | CNN | 45.880 | 2.945 | 37.129 | 0.334 | 7.104 | 2.627 | 0.440 | CSR | 27.853 | 1.834 | 21.168 | 0.472 | 6.701 | 2.968 | 0.326 | Proposed method | 47.024 | 4.259 | 37.523 | 0.500 | 7.487 | 2.461 | 0.606 | No. 8 | DTCWT_SR | 55.890 | 2.940 | 47.062 | 0.771 | 4.357 | 3.612 | 0.169 | NSCT_NSST | 52.610 | 4.872 | 47.647 | 0.868 | 4.832 | 3.906 | 0.380 | CNN | 87.236 | 6.208 | 73.907 | 0.721 | 5.688 | 3.933 | 0.294 | CSR | 54.208 | 4.753 | 47.214 | 0.794 | 4.245 | 3.383 | 0.285 | Proposed method | 92.152 | 15.279 | 76.870 | 0.943 | 4.333 | 4.947 | 0.505 |
|
查看原文
杨晓莉, 蔺素珍, 禄晓飞, 王丽芳, 李大威, 王斌. 基于生成对抗网络的多模态图像融合[J]. 激光与光电子学进展, 2019, 56(16): 161004. Xiaoli Yang, Suzhen Lin, Xiaofei Lu, Lifang Wang, Dawei Li, Bin Wang. Multimodal Image Fusion Based on Generative Adversarial Networks[J]. Laser & Optoelectronics Progress, 2019, 56(16): 161004.