方法比较

对比总结

1、09年的通过保边滤波解决 halo artifacts （halo artifacts are present at sharp edges where the luminance of the NIR and visible image are significantly different.）但是在亮度差异较大的区域，色彩失真比较大。（As we are only working with the luminance channel, colors might appear unrealistic in cases of extreme luminance changes 。
2、Feng et al.'s method fuses the RGB and NIR images based on transmission;
通过nir和RGB的不同求得transmission。但是这种方法的Feng etal.'s method creates ghost artefacts around cloud。
3.Dong 避免了 ghost artefacts around cloud，但是，由于仍采用了dark priority的方法，使得图像的颜色加重，不自然，看不清细节。视觉上提升效果不大。
4.Chang-Hwan Son的先给nir做color mapping，但是如果nir和RGB有slightly misaligned，或者 the luminance of the NIR and visible image are significantly different，就会有halo artifacts和color distortion。

image.png

chang-hwan的

image.png
我的
即使在加入了颜色和深度的正则化先验知识，进行约束，逼近先验。
但会出现边缘处的halo artifacts，并且颜色由于采用了dark priority的先验，仍然会不自然。
除了chang-hwan去雾效果好一些，细节多一些，其他的方法，融合后，图像的细节提升都不是特别显著，并且，近提升了远处的，在图像中，近处也会有一定的不清楚。我的能够对近处也有一定的提升作用，并且能够保持颜色的自然，视觉上能够达到一个比较好的效果。
image.png
image.png
image.png
image.png

1. 《COLOR IMAGE DEHAZING USING THE NEAR-INFRARED》Lex Schaul
   突破点：the halo artifacts that disappear by using edge-preserving filters.However, halo artifacts are present at sharp edges where the luminance of the NIR and visible image are significantly different. By making use of the WLS filters, these artifacts disappear.
   不足：As we are only working with the luminance channel, colors might appear unrealistic in cases of extreme luminance changes (Fig.3). We want to investigate in the future how our images can be improved by additional color processing.

   
   image.png
2. 《NEAR-INFRARED GUIDED COLOR IMAGE DEHAZING》Chen Feng
   Image dehazing in general involves two tasks, removing the airlight color effect and recovering the lost details.Removing the airlight color is fundamental to image dehazing. Inaccurate estimation of the airlight color could result in unwanted color shift
   in two-fold:
   Design an optimization framework to resolve the image de-hazing problem guided with NIR gradient constraints.
   Refine airlight color estimation by exploiting the differences between NIR and RGB channels.
   估计的大气光 都是一个固定值，但是大气是分布的，局部和全局是有差异的，并且对于t,在每个通道的t也是不同的。

3. 《Colour image dehazing using near-infrared fusion》Dong-Won Jang, Rae-Hong Park
   Schaul et al.'s method distorts the colour of the input image by fusing the NIR image. Although Feng et al.'s method removes the haze efficiently, some significant artefacts usually appear at cloud boundaries that are not aligned with those of the NIR image with small transmission values.
   Regions or images without haze remain unaltered. Therefore, our method can be applied whether haze is actually present or not.
   Both methods efficiently remove haze in the mountainous area beyond the stretch of water. However, the input images are slightly misaligned, and Feng et al.'s method fuses the RGB and NIR images based on transmission; as expected, Feng etal.'s method creates ghost artefacts around cloud, as shown in Fig. 6c. In contrast, the proposed method fuses images based on the HF distribution of the local patch and so, as shown in Fig. 6d, no ghost artefact is observed in misaligned regions.

   
   image.png
   

   4.There are serious discrepancies in terms of brightness and image structures between the near-infrared image and the visible color image. Due to this discrepancy, the direct use of the near-infrared image for haze removal causes a color distortion problem during near-infrared fusion. The key objective for the near-infrared fusion is therefore to remove the color distortion as well as the haze.
   In short, the new issue for the near-infrared fusion is to preserve high visibility of the captured near-infrared image and to remove color distortion that appears during image fusion
   加入了颜色和深度的正则化先验知识，进行约束。逼近先验。

考虑到的现象还有实验的一些发现

NIR的不同物质区域的反射特性不同。

物质对近红外和RGB四通道的反射率差异比较大，通过NVDI指数，植被等区域近红外强于RGB，天空雾气水等近红外弱于RGB。就是图像的对比度差异比较大。RGB三个通道和

近红外图像不能理解为一个单纯的亮度图像。

然而由于物体成像根据Retinx理论分为入射光和反射光，所以，当想要求RGB和NIR的反射的差异时，先使用了同态滤波去除了入射光的影响。

这个差异图，代表了自然场景下，物质对NIR和RGB的反射率的不同，是考虑到了物质的反射的物理特性。------------------

引导滤波和双边滤波的效果类似，但是，速度快。

因为 NIR和RGB的对比度是不同的，所以在将detail层融合的时候，尽量不引进近红外的对比度信息，而是主要引进的是纹理信息。而通过商图像(光照不变的 illumination invariant)，可以最大限度的隐藏对比度信息，保留纹理信息。

The ratio is computed on each RGB channel separately and is independent of the signal magnitude and surface reflectance. The ratio captures the local detail variation in and is commonly called a quotient image [Shashua and Riklin-Raviv 2001] or ratio image [Liu et al. 2001] in computer vision

We use the detail information from the NIR image to both reduce noise in the RGB image and sharpen its detail. The base layer of RGB image contains low luminance information as perceived by humans visual system, thus the NIR base layer is discarded. 如此，可以保持融合以后颜色的自然。