Description
Your task this week is to implement a
few basic image processing techniques
on arrays of pixels. Specifically, you
must perform conversion from RGB
(red, green, and blue) pixels to HSL
(hue, saturation, and luma/luminance)
and back, edge detection using
convolution with Sobel kernels, and
histogram equalization to balance the
distribution of brightness across the
image.
Code provided to you makes use of
your functions to process PNG image
files. Starting with the image in the upper left, for example, edge detection produces the image shown in
the lower left. Sharpening those edges produces the image in the upper right, and following up with
histogram equalization produces the image in the lower right.
The objective for this week is for you to gain experience with using multi-dimensional arrays to represent
information in C.
Background
Image Representations: Images are often stored and represented using red, green, and blue (RGB)
intensities, which map naturally into displays. For example, image data stored as a Portable Network
Graphics (PNG) file, the pixels are represented in this way. When we want to process those images,
however, other representations are useful. The HSL (hue, saturation, luma) representation was developed
by computer graphics researchers to better match human perception. The luma component represents
luminosity, also known as intensity, and corresponds roughly to the brightness of a given pixel. Hue
represents the color of the pixel, and saturation represents how much of that color is present in the pixel (a
saturation of 0 is white, for example).
For this assignment, you must write C functions that convert between the RGB and HSL representations
and back. Specifically, for each pixel in an image, its RGB representation is stored in a fairly common way
as three separate bytes, with 8 bits of red (0-255), 8 bits of green (0-255), and 8 bits of blue (0-255). The
HSL representation is defined for the purposes of this assignment: while it is similar conceptually to those
that you can read about, the details do not match precisely. To maintain more information about the original
RGB color, we make use of 16-bit unsigned values for H, S, and L. All equations below correspond to
the use of integer arithmetic, so be careful about the order of operations.
To obtain H, S, and L from RGB, we start by finding the minimum and maximum RGB values, as follows:
𝑀 = max (𝑅, 𝐺, 𝐵) 𝑁 = min (𝑅, 𝐺, 𝐵)
The luma L is then given by the sum of M and N, and we compute the chromaticity C as the difference
between M and N (C is used to compute H and S):
𝐿 = 𝑀 + 𝑁 𝐶 = 𝑀 − 𝑁
Using C, we can then find the pixel’s saturation S:
𝑆 =
⎩
⎪
⎨
⎪
⎧
0 if 𝑀 = 0 or 𝑁 = 255
0x8000 𝐶
𝐿
if 0 < 𝐿 ≤ 255 0x8000 𝐶 510 − 𝐿 otherwise Be sure that your implementation obeys the order of operations in the equations above (multiplication before division) to avoid differences in integer arithmetic results. Finally, we compute H using the following equation: 𝐻 = ⎩ ⎪ ⎪ ⎪ ⎨ ⎪ ⎪ ⎪ ⎧ 0 if 𝐶 = 0 0x2000 (𝐶 + 𝐺 − 𝐵) 𝐶 if 𝐶 > 0 and 𝑀 = 𝑅
0x2000 (3𝐶 + 𝐵 − 𝑅)
𝐶
if 𝐶 > 0 and 𝑀 > 𝑅 and 𝑀 = 𝐺
0x2000 (5𝐶 + 𝑅 − 𝐺)
𝐶
otherwise
Again, be sure to perform the division last in each case, and to use large enough integers to avoid overflow
in your computations.
The meaning of hue is somewhat arbitrary, and is related to the mixing of display color components. The
definition above is also rotated for simplicity of calculation, and thus does not match any standard
definitions. For interest, the colors magenta, red, yellow, green, cyan, and blue correspond to H values
0x0000, 0x2000, 0x4000, 0x6000, 0x8000, and 0xA000, respectively.
Converting from HSL back to RGB is slightly more complicated, but is also done on a per-pixel basis.
Again, your code must use the definitions and equations given in this specification.
We start by calculating chromaticity C from S and L as follows:
𝐶 =
⎩
⎪
⎨
⎪
⎧
𝑆 𝐿
0x8000 if 𝐿 ≤ 255
𝑆 (510 − 𝐿)
0x8000 otherwise
As before, be careful about operation order. The minimum RGB component N and maximum RGB
component M can then be recovered from C and L:
𝑁 = (𝐿 − 𝐶)/2 𝑀 = 𝑁 + 𝐶
Then we need to break apart the distinct regions of hue H. To do so, we define a major and a minor part of
the hue as follows:
𝐻୫ୟ୨୭୰ =
ு
௫ଶ
𝐻୫୧୬୭୰ = 𝐻 & 0x3FFF
Remember that both values are integers. For simplicity, we have used the C bitwise AND operator in the
definition of 𝐻୫୧୬୭୰.
The third RGB component, which we call T, can be found using 𝐻୫୧୬୭୰ as follows:
𝑇 =
⎩
⎪
⎨
⎪
⎧
𝑁 +
𝐶 (𝐻୫୧୬୭୰ − 0x2000)
0x2000 𝐻୫୧୬୭୰ ≥ 0x2000
𝑁 +
𝐶 (0x2000 − 𝐻୫୧୬୭୰)
0x2000 otherwise
The major hue ranges from 0 to 5, and the order of RGB components depends on this value as follows:
(𝑅, 𝐺, 𝐵) =
⎩
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎧
(𝑀, 𝑁, 𝑇) if 𝐻୫ୟ୨୭୰ = 0
(𝑀, 𝑇, 𝑁) if 𝐻୫ୟ୨୭୰ = 1
(𝑇, 𝑀, 𝑁) if 𝐻୫ୟ୨୭୰ = 2
(𝑁, 𝑀, 𝑇) if 𝐻୫ୟ୨୭୰ = 3
(𝑁, 𝑇, 𝑀) if 𝐻୫ୟ୨୭୰ = 4
(𝑇, 𝑁, 𝑀) otherwise
Edge Detection: Once we have an image in an HSL representation, we can use the L channel (the luma
values for the pixels) to identify edges in the image. Edge detection is useful for image segmentation and
computer vision applications, and the kinds of operations that you will implement often form the first step
in image processing pipelines for any type of recognition. The mathematical operations, specifically
convolution, are also used widely in implementing convolutional neural network layers, an important
component in most AI/machine learning applications.
For this assignment, we make use of Sobel edge detection, which uses two 3×3 kernels (matrices) to identify
changes in intensity in the horizontal and vertical directions in an image. The two kernels are shown here:
𝐾ଡ଼ =
1 0 −1
2 0 −2
1 0 −1
൩ 𝐾ଢ଼ =
1 2 1
0 0 0
−1 −2 −1
൩
In the kernels, X values increase to the right, and Y values increase as one moves down. These axes
directions match those commonly used with images.
To perform edge detection, we convolve the kernels with the image.
Consider the horizontal kernel, 𝐾ଡ଼. For each pixel in the image, we
logically align the center value (the blue one) in the kernel with the
pixel, multiply each kernel value by the corresponding image luma
(one pixel’s L value), and sum up the products. For example, to
compute the value of the convolution of 𝐾ଡ଼ with the blue pixel
highlighted in the example L data to the right, we multiply (from the
first row) 9 by 1, 8 by 0, and 4 by -1—these sum to 5. From the
second row, we multiply 2 by 2, 3 by 0, and 4 by -2—these sum to -4.
And, from the third row, we multiply 7 by 1, 8 by 0, and 7 by -1—
these sum to 0. The total sum is thus 1, which is the result of the convolution for that pixel.
Denoting the two-dimensional convolution operation by *, we compute 𝐺ଡ଼ = 𝐾ଡ଼ ∗ 𝐿 and 𝐺ଢ଼ = 𝐾ଢ଼ ∗ 𝐿.
These values can then be used to find the magnitude and direction of the gradient (the rate of change) at
each point in the image. You might wonder what we should do at the image boundaries, where some
values in the kernel do not correspond to pixels (choose any edge pixel as the blue pixel in L). For
simplicity, we do not compute anything for those pixels.
4 6 3 7 5 9
4 7 9 8 4 3 1 0 -1
9 2 2 3 4 9 2 0 -2
7 8 7 8 7 7 1 0 -1
8 7 8 7 2 3
4 5 3 4 9 3
Histogram Equalization: Operations to balance color and lighting in an image are routinely performed
by modern digital cameras. You may have noticed, for example, how much better each new smartphone
captures images in twilight than did the last generation. For the last operation in this assignment, we
make use of a simple technique to evenly distribute the luminosity of pixels in an image.
Starting again with the L channel for an image with P pixels, we begin by computing a histogram of L
values over the entire image. In this assignment, our HSL definition restricts L values to the range
[0,510], so the histogram is fairly small. Let’s call that histogram H, then compute the cumulative
histogram K as follows:
𝐾(𝑖) = 𝐻(𝑗)
ୀ
If the pixels have a uniform distribution of luma values, we expect to find
𝐻(𝑖) =
ହଵଵ
and 𝐾(𝑖) =
(ାଵ)
ହଵଵ
However, most images are not so uniform. Fortunately, we can make the lighting more uniform by
remapping each of the 511 luma values to one that produces a more uniform distribution. Given a
specific luma value X, K(X) corresponds to the number of pixels with luma value at or below that level,
which we want to be equal to (X+1) L / 511. To make lighting uniform, let’s choose another value, X’,
such that
𝐾(𝑋) =
൫
ᇲା ଵ൯
ହଵଵ
.
Solving for X’, we obtain
𝑋
ᇱ =
511 𝐾(𝑋)
𝐿
− 1
In other words, we should replace the luma value for every pixel that has L=X with the new value X’, as
given above. Somewhat arbitrarily, but following convention, we choose to round the division up rather
than down, and, since the numbers may overflow 32 bits for large images, we make use of 64-bit values
using the following, more C-like expression:
𝑋
ᇱ = (511 ∗ (int64_t)𝐾(𝑋) + 𝐿 − 1) / 𝐿 − 1
By replacing all luma values in the image, we make lighting more uniform.
Pieces
Your program will consist of a total of three files:
mp6.h This header file provides type definitions, function declarations, and brief
descriptions of the subroutines that you must write for this assignment.
mp6.c The main source file for your code. A version has been provided to you with full
headers for each of your subroutines; be sure to read the function headers
before you begin coding. You need merely fill in the body of each subroutine.
A third source file is also provided to you:
main.c A source file that interprets commands, reads and writes PNG files, calls your
subroutines, and sharpens edges in an image. You need not read this file, although
you are welcome to do so.
A Makefile and several sample images have also been provided for convenience. The Makefile allows
you to build the mp6 executable by simply typing make at the command line. The images are in a
subdirectory. We have also included a compiled version of a solution, called gold, that you can use to
generate additional results for comparison with your own solution. The outputs of your program should
match the gold version exactly in all cases.
The Task
You must write four C subroutines in this assignment. Don’t panic: the total amount of code needed in my
version was only about 130 extra lines! I recommend implementing the functions in the order described in
this document, which allows you to perform some amount of debugging as you implement rather than trying
to debug everything at once.
Step 1: Implement RGB to HSL conversion.
Converting RGB to HSL is slightly simpler than the reverse conversion. Implement the C function
void convert_RGB_to_HSL (int32_t height, int32_t width,
const uint8_t* red, const uint8_t* green, const uint8_t* blue,
uint16_t* H, uint16_t* S, uint16_t* L);
in mp6.c. Use the equations presented earlier in this specification. Each of the six arrays is a twodimensional image laid out in row-major order in memory. Be sure that you understand the discussion in
lecture and know how to compute the linear index for any (x,y) coordinate in the image. Note that the
image width (and height) are given to you as parameters.
Once you have finished this function, compile the program and use the HSL output option to make sure
that you have produced the correct answers. Note that the output file in this case is human-readable, and is
not a PNG image. You can use the sample images provided as inputs or use your own (in PNG format),
and can use the gold executable to produce correct answers for any input image.
Step 2: Implement HSL to RGB conversion.
Next, implement the reverse conversion by completing the C function
void convert_HSL_to_RGB (int32_t height, int32_t width,
const uint16_t* H, const uint16_t* S, const uint16_t* L,
uint8_t* red, uint8_t* green, uint8_t* blue);
in mp6.c. As with the first step, use the equations presented earlier in this specification. To keep the
computations reasonably simple, the equations are slightly lossy, so converting RGB to HSL and back does
not always produce the original RGB value. Feel free to improve the design on your own if you are
interested, but be sure to turn in a solution that matches this specification exactly.
Again, each of the six arrays is a two-dimensional image laid out in row-major order in memory.
Once you have finished this function, compile the program and use the RGB→HSL→RGB output option
to make sure that you have produced the correct answers. This option (and all others except the first option)
produces a PNG image as output. Visual comparison of the original and final images should be enough for
initial testing, but you can also use image comparison tools to compare your program’s output with that of
the gold executable provided to you.
Step 3: Implement Sobel edge detection.
You are now ready to implement Sobel edge detection on the L channel. To do so, complete the C function
void compute_sobel_kernels (int32_t height, int32_t width,
const uint16_t* L, int32_t* Gx, int32_t* Gy);
in mp6.c. Sobel edge detection is fairly standard, but we still suggest following the specification given
earlier, as both gradient direction and handling of boundaries is somewhat arbitrary, and yours must match
ours exactly for credit.
As with the earlier functions, each of the three arrays is a two-dimensional image laid out in row-major
order in memory. Although you need not produce any values for the boundary pixels in the Gx and Gy
output arrays, these boundaries are included in the arrays. In other words, all three arrays have size
height * width. The boundary pixels will be ignored—your code need write nothing into them.
Once you have finished this function, compile the program and use the edge detection output option to
make sure that you have produced the correct answers. This option produces a PNG image as output.
Visual comparison of your output with that of gold should be enough for initial testing, but a more thorough
comparison with image comparison tools is recommended.
The edge detection output uses only the magnitude of the gradient (the Gx and Gy arrays). You can check
that you didn’t flip the direction by using the edge sharpening output option, which also uses the gradient
direction to sharpen the edges in an image. Note that the code for doing so has been provided to you, so
you need do nothing other than run the program with a different option to check that your program’s output
matches that of gold.
Step 4: Implement histogram equalization.
Finally, implement the C function
void equalize_intensities (int32_t height, int32_t width, uint16_t* L);
in mp6.c. Here, again, the implementation is fairly specific to our HSL definition, so follow the description
earlier in this specification exactly, and be sure to use the final expression given for re-mapping intensities.
For full credit, build a lookup table for intensity remapping rather than re-calculating the new
intensity for each pixel.
Only the L channel array is provided to this function, both as input and as output. The image height and
width are also given, of course.
Once you have finished this function, compile the program and use the all + equalization output option to
make sure that you have produced the correct answers. This option produces a PNG image as output.
Visual comparison of your output with that of gold should be enough for initial testing, but a more thorough
comparison with image comparison tools is again recommended.
Specifics
Be sure that you have read the function headers and other information in the code before you begin coding.
Your code must be written in C and must be contained in the mp6.c file in the mp/mp6 subdirectory
of your repository. We will NOT grade any other files. Changes made to any other files WILL
BE IGNORED during grading. If your code does not work properly without such changes, you
are likely to receive 0 credit.
You must implement the convert_RGB_to_HSL, convert_HSL_to_RGB,
compute_sobel_kernels, and equalize_intensities functions correctly.
You may assume that all parameter values are valid when your functions are called, provided that
you also produce only valid parameters in your own functions (incorrect values produced by your
functions may be passed back to your other functions). In particular,
o arrays will be sized as specified and laid out in row-major order,
o red, green, and blue values will be in the range 0-255, and
o luma values will be in the range 0-510.
We will not test on images larger than a few million pixels, although your code should not have
any explicit checks on the provided values of image height and width.
Your routines’ outputs must be correct.
Do not make any assumptions about the order or number of times that your functions are called.
For testing, we will make use of additional code, and you will lose points if your functions do not
work repeatedly as specified, whether or not other functions are called before or after them.
Your code must be well-commented. Follow the commenting style of the code examples provided
in class and in the textbook.
Compiling and Executing Your Program
When you are ready to compile, type
make
The Makefile describes the dependences between your source files and the mp6 executable, so whenever
you make changes to a source, the command above should automatically rebuild your code. The “-g”
argument is included to tell the compiler to include debugging information so that you can use gdb to find
your bugs (you will have some).
The “-Wall” argument is also included to tell the compiler to give you warning messages for any code that
it thinks likely to be a bug. Track down and fix all such issues, as they are usually bugs. Also note that if
your code generates any warnings, you will lose points.
If compilation succeeds, you can then execute the program by typing, “./mp6” (no quotes) to get
instructions on how to use the program, including how to specify input file, output file, and choice of
processing option.
Remember that testing is your responsibility. Follow the suggested ordering and guidelines for
developing your code and comparing with the output of the gold program.
Grading Rubric
Functionality (70%)
20% – convert_RGB_to_HSL function works correctly
20% – convert_HSL_to_RGB function works correctly
15% – compute_sobel_kernels function works correctly
15% – equalize_intensities function works correctly
Style (15%)
10% – conversion code (both convert_RGB_to_HSL and convert_HSL_to_RGB) uses a
variable to store pixel index rather than re-computing linear array indices on every array access
5% – equalize_intensities uses a lookup table rather than re-computing a new intensity
value for each pixel
Comments, Clarity, and Write-up (15%)
5% – introductory paragraph explaining what you did (even if it’s just the required work)
10% – code is clear and well-commented, and compilation generates no warnings (note: any
warning means 0 points here)
Note that some categories in the rubric may depend on other categories and/or criteria. For example, if you
code does not compile, you will receive no functionality points. As always, your functions must be able to
be called many times and produce the correct results, so we suggest that you avoid using any static storage
(or you may lose most/all of your functionality points).
