This activity is pretty simple, you have an image with unequal bright and
dark tones, and then you try to equalize it to make it look better. I generally
enjoyed doing this activity especially since it brings satisfaction from the revelation of what a poorly contrasted image was hiding.
In the intro part of the manual, Ma'am Jing said that we should not
trust our eyes. I say, yes it is true. Colors of images depend not just on its
hue but mostly on the camera settings (trinity of colors, remember? 187?) Good
for us, humans, our eyes have rods that adjusts contrasts on the environment
that we are looking at so we can have a good perception of the reality. Due to
the vast improvement of science, most the technology of today can mimic or surpass the way our eyes work. Even some of our hand held phone
cameras can already capture information that our eyes cannot see.
There will always be cases however, that some important data, such as
crime scene evidence, extinct animal sightings, underwater explorations, and
many others, may sometimes be poorly contrasted due to the intense environment
settings (dark room, too bright background, camouflaging objects, moving objects, etc) . With the proper tools, the stored information in those images can be
extracted to maximize the value of those seemingly lost data, and reveal things
our eyes have never seen before.
For the activity, the PDF of the image was obtained. Naturally, for a
good image, the PDF should somehow look like a horizontal line, indicating equal distribution of all the possible intensities. A too dark image would
have peaks at the left part of the PDF, which are the 0 or the dark pixels and
the PDF may cluster at the left part. A too light image will have peaks or
clusters at the right part, indicating 255 or the white pixels. Some image may
have intense contrast such that it has peaks far apart from each other, which
are too light and too dark. A poorly contrasted image however, has a PDF
clustered at one side. A highly contrasted image has peaks at the opposite end
of the graph showing too dark and too light pixels. To enhance the image, the
desired PDF should be a horizontal line, spread out along all intensities. This
is the expected PDF of a well-contrasted image. The image enhancement by
histogram manipulation basically is using a desired CDF (or the cumulative
distribution function of the PDF) and back projecting the new grayscale values
of the image referencing on the original CDF. So, to equalize the
intensities, the desired CDF will be a line with a positive slope.
I have to apply the histogram manipulation to a poorly contrasted image.
It took me a hard time to search for an image in my phone since the images nowadays are
of good quality and I've already deleted unnecessary photos. I stumbled upon
this picture of Rey I've taken when IPL was going for lunch after a Gen Clean
in the lab. The image has a bad contrast since the focus was set on the sky at the
back ground, which made the foreground look very dark. You wouldn't know it was Rey
in the picture at first, right? I even questioned the reason myself on why I
still haven't deleted it since I can't comprehend anything from the original
image. After swiping some other pictures taken a few moments after the photo
was taken (one will be shown near the end of this blog post), then I realized
what this photo was supposed to show. I still can't remember though, why I let
this bad photo consume space on my phone.
Anyway, going straight to the point, I used the original image as shown
in figure 1(a) and tried to equalize the histogram of the intensities. The
result as shown in figure 1(b) is pretty okay. However it became lighter and
looks faded as confirmed by the low contrast shown by the histogram. Also, a
big spike at some pixel lighter than the original peak in the histogram was
created. Due to the limitation of the back projection done in the methodology,
most of the domineering pixels, which are dark pixels, in the original image
got transformed to a lighter pixel as shown by the new CDF in figure 1(d) where
the values at the start of the line has 0 y-values. This result seems like the
image did not get enhanced but was rendered far from the original image.
However, a lot of details was revealed which are not observable in the original
image.
Figure
1. An enhanced image originally taken with a focus on the sky. (a) The original
colored image. (b) The image enhanced by linearizing the intensities. (c) The
distributions of the intensities for the original and enhanced image. (d) The
CDF
Since the image in figure 1(a) was captured against the light, where the
background is a very bright light source, linearizing the contrast of the image
makes the image faded due to the limitation or the discreteness of the CDF
felt during the backprojection. The rendering of the image was improved by
using the fraction of the original intensities after linearizing the
intensities of the image. The result shown in figure 2 (a) shows an improved image. The PDF of the
histogram also shows that the peak shifted to the left which is nearer to the
original peak in the histogram. Not much change in the number of revealed
details was observed compared to the result using the linear CDF.
Figure
2. (a) The image enhanced by linearizing the intensities and using a fraction
of the original intensities to render a better image (b) The histogram of the
intensities showing the peak nearer to the peak of the original image. (c) The
CDF showing a slight distortion due to the original CDF of the image.
Discovering the effect of other desired CDFs, I used the sigmoid, exponential
and log functions. The sigmoid CDF resulted to an image showing more details
than the original. The image however is more faded that the image obtained from
the linear CDF. The small tree branches are not well defined with the sigmoid
CDF compared to the linear CDF. The exponential CDF in figure 4 rendered a bad
image as a lot of the fine details at the back ground were removed. Some
details in the foreground however was revealed but the image is so faded. The
result using the log CDF is okay since the rendered image is not faded and a
lot of details were revealed.
Figure
3. (a) The image enhanced using a sigmoid CDF. (b) The histogram showing the
dark colored pixels are removed and the peak shifted to a lighter pixel. (c)
The CDF of a sigmoid.
Figure
4. (a) The image enhanced using a CDF where x = ya where a is a positive
number. (b) The histogram showing increased height of the peaks near light
pixels. (c) The CDF resulting from the discretization of the original CDF.
Figure
5. (a) The image resulting from a desired CDF where x =
ya where a is a
fraction. (b) The histogram showing the peaks at the dark colored pixels
shifted to the right compared to the original histogram. (c) The CDF of the
image.
So far, with the original image, the best enhancement was made using the
linear CDF and the log CDF. Yes, it is true that a lot of details not seen in
the original was revealed in the enhanced images. However, the enhanced images
are only in its intensity. The RGB details of the images were already lost. I
was challenged on thinking of a way to be able to render a colored version of
the enhanced image. First, I thought of manipulating the CDF of the R, G, and B
channels separately and then restacking the separate channels to show the
original image. Figure 6 (a) shows the rendered image after separately
linearizing the CDF of the channels, and figure 6 (b) shows the result after
using a log CDF of the channels. The images rendered is not that good but is actually
still better than the original image. A lot of details on the color of the
objects in the image which was not very visible in the original image was now
revealed such as the blue shirt worn by RM (the person at the left side of
Rey), which looks black in the original image in figure 1 (a). In the resulting
images, the R, G, and B ratios were not preserved, which explains the faded
result since most of the colors became white due to the linearized intensity of
the R, G, and B channels. Figure 6 (c) shows the resulting image while
preserving the ratios among the color channels with reference white balancing.
As can be seen, red noise is visible and the skin color is reddish. I also
tried using the hsv channels of the images however I failed to reconstruct the
ideal colors of the objects since the image became yellow-greenish all over.
After doing a lot of trial and error with using the RGB or the hsv channels and
using difference desired CDFs, I accounted the red noise to the lost data due to
the poor contrast in the original image.
To compare the result to the expected real world colors, I showed figure
6 (d) which is the image I captured a few seconds after capturing figure 1 (a)
but now with the smart phone camera adjusted to focus on the foreground. See
how much details was lost by comparing figure 1 (a) and figure 6 (d). It is
indeed vital to know how to focus on the important things to be able to show
the essential details. Good thing, we have technology to extract the details we
lost due to are mistakes. We are still learning anyway.
Figure 6.
(a) Image rendered by linearizing the CDF of the R, G, and B channels. (b) The
image rendered be separating the R, G, and B channels and using a log CDF for
each. (c) Image rendered using a linear CDF for each R, G, and B channels and
preserving the ratios among each color channels with reference white balancing.
(d) A colored image taken a few moments after the image in figure 1 (a) was
taken, but with the proper camera settings to show the details in the
foreground.
For this activity, I give myself 10/10 since I was
able to enhance the intensity of a poorly contrasted image using different
CDFs. I had a lot of fun doing the activity but it took me a long time to
figure out how to properly preserve the color after the histogram manipulation.
For the image in figure 1 (a) the best image was shown in figure 6 (c) which is
not that good. I tried using other images and the results are better.
For the image is shown in figure 7. The left panels
shows the original image (figure 7 (a) and (d)), the middle panels shows the
images with the contrast enhanced using a linear CDF (figure 7 (b) and (e)),
and the right panels shows the images with the histogram of the RGB channels
separately manipulated. Figure 7 (a) is an image of my brother with our dog,
the image is a bit highly contrasted since it was taken against the light
source, making the skin look dark and
the window railings at the lower left part unseen. After the manipulation shown
in figure 7 (c), the edges of the window railings were revealed and the skin
became lighter. The edge of the roof at the background was also revealed.
Figure
7 (d) is an image of the BS5 CSG boys during SPP. Due to the bad camera
settings, a dark void can be seen near Justin’s hair and at the left side of
Rey’s head. The general mood of the image is also cold and a bit gloomy. After
the manipulation as shown in figure 7 (f), the edges at the stage (near
Justin’s hair) were revealed. Details were also shown near Rey’s hair. Due to
the manipulation, the surroundings became lighter giving the image a joyful
atmosphere.
Figure 7. (a) and (d) The original colored images. (b) The enhanced
image of (a) using a linear CDF. (c) The image in (a) enhanced with the color information
preserved. (e) The enhanced image of (d) by linearizing the CDF. (f) The image
in (d) enhanced with the color information preserved.
Reality is to be found in lightness and darkness - Pablo Picasso
