1. 论文总述
本文是2007年的一篇文章,也是图像融合中很经典的算法,对几张需要融合的图像,逐像素计算权重,然后根据权重进行逐像素融合;图像直接融合容易出现难看的比较假的缝隙边缘或者光晕等,于是作者将原图分解为拉普拉斯金字塔、将权重图分解为高斯金字塔,然后将这两个金字塔相乘,最后恢复出原图。
优势:与HDR相比,不需要相机标定等参数,还跳过了tone mapping这步,而且HDR融合的图有点假,没有Exposure Fusion的还原度高。
缺点: 动态范围不高,但是输出的图可以直接在显示设备上进行display,不需要tone_mapping?
HDR与Exposure Fusion的对比可以参考这里的讨论
注: 权重的具体计算可以利用:对比度、饱和度、Well-exposedness(作者用的),当然也可以根据需求自己设计权重指标。
In this paper, we propose to skip the step of computing a
high dynamic range image, and immediately fuse the multiple exposures into a high-quality, low dynamic range image,
ready for display (like a tone-mapped picture). We call this
process exposure fusion; see Fig. 1. The idea behind our approach is that we compute a perceptual quality measure for
each pixel in the multi-exposure sequence, which encodes
desirable qualities, like saturation and contrast. Guided by
our quality measures, we select the “good” pixels from the
sequence and combine them into the final result.
Exposure fusion has several advantages. First of all,
the acquisition pipeline is simplified, no in-between HDR
image needs to be computed. Since our technique is not
physically-based, we do not need to worry about calibration of the camera response curve, and keeping track of
each photograph’s exposure time. We can even add a flash
image to the sequence to enrich the result with additional
detail. Our approach merely relies on simple quality measures, like saturation and contrast, which prove to be very
effective. Also, results can be computed at near-interactive
rates, as our technique mostly relies a pyramidal image decomposition. On the downside, we cannot extend the dynamic range of the original pictures, but instead we directly
produce a well-exposed image for display purposes.
2. Quality Measures
3. 权重图和原图直接相乘融合的缺点
with Ik the k-th input image in the sequence. Unfortunately, just applying Eq. 1 produces an unsatisfactory result. Wherever weights vary quickly, disturbing seams will
appear (Fig. 4b). This happens because the images we are
combining, contain different absolute intensities due to their
different exposure times. We could avoid sharp weight map
transitions by smoothing the weight map with a Gaussian
filter, but this results in undesirable halos around edges, and
spills information across object boundaries (Fig. 4c). An
edge-aware smoothing operation using the cross-bilateral
filter seems like a better alternative [22, 9]. However, it is
unclear how to define the control image, which would tell
us where the smoothing should be stopped. Using the original grayscale image as control image does not work well,
as demonstrated in Fig. 4d. Also, it is hard to find good parameters for the cross-bilateral filter (i.e., for controlling the
spatial and intensity influence).
所以才有了本文提出的原图拉普拉斯金字塔分解、权重图高斯金字塔分解
4. 多分辨率融合效果好的背后原因
Multiresolution blending is quite effective at avoiding
seams (Fig. 4), because it blends image features instead of
intensities. Since the blending equation (1) is computed at
each scale separately, sharp transitions in the weight map
can only affect sharp transitions appear in the original images (e.g. edges). Conversely, flat regions in the original
images will always have negligible coefficient magnitude,
and are thus not affected by possibly sharp variations in the
weight function, even though the absolute intensities among
the inputs could be different there.
Instead, we blend the pyramid coefficients based on a scalar
weight map, but do not directly process individual coeffi-
cients at different levels. Measures like saturation and wellexposedness are hard to evaluate directly from pyramid coefficients. Our technique basically decouples the weighting
from the actual pyramid contents, which enables us to more
easily define quality measures. In fact, any measure that can
be computed per-pixel, or perhaps in a very small neighborhood, is applicable.
5. 本文方法的不足之处
In Fig. 6, our result
contains a spurious low frequency brightness change, which
is not present in the original image set. It is caused by a
highly varying change in brightness among the different exposures. Intuitively speaking, this artifact can be considered
as a very blurred version of the seam problem, illustrated
in Fig. 4b. Constructing a higher Laplacian pyramid partially solves this problem. However, the pyramid height is
also limited by the size of the downsampling/upsampling
filter [3].