Goals

Overview

You will create binarize(…), a function that works with your code from HW11 to convert an image to monochrome (black & white) using the “adaptive threshold” method.

Adaptive Thresholding

Thresholding is a technique used to convert an image to monochrome (black & white). It is important for problems such as scanned document processing, OCR, image segmentation, and medical imaging. Pixels brighter than some threshold are set to white. The rest are set to black. The key question becomes this: What should the threshold be?

If the image is color, you will need to calculate what each pixel would be as grayscale. In the image below (taken with a smartphone), the image at left is actually color. Notice that there is a little tint from the lighting in the room. The image right is true grayscale. Your code should be able to handle either one.

color

gray

In the simplest case, you could use 50% gray ( ) so that        would be converted to       . That method—often called fixed threshold because it uses the same threshold value for all pixels—generally works well for scanned documents, as long as the brightness and contrast were set properly when they were scanned. However, in cases of photographs taken in uneven lighting or certain medical applications, a fixed threshold provides poor results, as you see below.

gray	monochrome, fixed threshold

Even if we were to use 30% gray or 70% gray, we would still have similar problems. Because the lighting is different in different regions of the image, there is no one good threshold to use for every pixel. The solution is adaptive threshold. The basic idea is that for each pixel in an image, you may use a different threshold value to decide whether that pixel should be black or white.

gray	monochrome, adaptive threshold

There are many good ways to decide the threshold, none of them perfect. For this assignment, we will keep it simple:

For a neighborhood radius r, a pixel at (x₀, y₀) shall be white if and only if its intensity is greater than the average intensity of all pixels in the (2r + 1) × (2r + 1) neighborhood surrounding that pixel. That neighborhood shall consist of all pixels at (x_i, y_i), such that |x_i - x₀| ≤ r and |y_i - y₀| ≤ r, including the current pixel being thresholded. Pixels that are not white shall be black. In calculating the threshold, do not include pixels that are beyond the boundaries of the image. This means that for such edge pixels, the neighborhood will be smaller.

Example: Suppose we binarize the following 6x6 image with radius=1.

The pixel at (x=2, y=3) currently has intensity value of 80. The neighborhood (radius=1) consists of pixels having intensity values 45, 50, 55, 75, 80, 85, 145, 135, and 125. The average of those is 88⅓. Since 80 ≤ 88⅓, the pixel at (x=2, y=3) shall be black.

Always calculate the threshold based on only the values in the input image.

(Optional) More information about adaptive threshold:

• Adaptive Threshold (Prof. Robert Fisher, University of Edinburgh)

Multi-threaded programming

Calculating adaptive threshold can be computationally intensive, especially on high resolution images. Therefore, you will implement the adaptive threshold algorithm in parallel by using the pthread library. How can parallelism be implemented for this algorithm? Each pixel in the output image must have its threshold calculated and its intensity value set. Threads can be created such that each thread operates on a particular region of pixels in the image. How you choose to allocate pixels to a particular thread is up to you!

Your binarize(…) will create a copy, which makes your life a lot easier than it would be if they were modifying the image in place.

A thread is defined as an independent stream of instructions that can be scheduled to run by the operating system. Multi-threading allows multiple threads to exist within the context of a single process. These threads share a process' resources but are able to execute simultaneously and/or independently. You are responsible for creation of num_threads threads and their management. See the man pages for pthread_create(…) and pthread_join(…) for details on their parameters and how to pass data to your worker function. You can refer to example#1 and example#2 covered in class. Also for more examples of thread management you can refer here.

Warm-up exercises

This assignment includes a warm-up exercise to help you get ready. This accounts for 10% of your score for HW13. Scoring will be relatively light, but the usual base requirements apply.

Zero an array. Create a function bool zero array(int✶ array, int array length, int num threads, const char✶✶ a error) that sets every element of the given array to zero using num_threads threads.

Opt out.

In a hurry, and don't need the practice? This warm-up is here to help you learn what you need to succeed on the rest of this assignment—not to add additional work. Therefore, we give you an option. Those who feel that they do not need the practice may "opt out". by modifying warmup.c so that it does nothing but print the following message exactly and then exit:

I already know this and do not need to practice.

If you do that, then your score for HW13 will be based solely on the rest of this assignment. If you leave the warmup.c undone, if you do not turn in a warmup.c, or if the message it prints does not match perfectly, then you will receive 0 for the warmup portion of this assignment (10%).

Start

To get the starter files, enter 264get hw13 in bash. You will get the header files (bmp.h and mtat.h), a skeleton warmup.c, and several image files. The image files with "_bw_" in the filename are the output of the binarize(…) function, using the given radius.

You are strongly encouraged to work with the smallest image (6x6), using only 1, 2, or 3 threads. Keeping things simple and small will make your development and debugging more manageable.

Requirements

Compile

gcc -pthread -o test_mtat mtat.c bmp.c test_mtat.c

Pre-tester ●

The pre-tester for HW13 has not yet been released. As soon as it is ready, this note will be changed.

Submit

In general, to submit any assignment for this course, you will use the following command:

264submit ASSIGNMENT FILES…

For HW13, you will type 264submit hw13 warmup.c mtat.c mtat.h bmp.c bmp.h test_mtat.c from inside your hw13 directory.

You can submit as often as you want, even if you are not finished with the assignment. That saves a backup copy which we can retrieve for you if you ever have a problem.

How much work is this?

This assignment does not require a lot of code. However, writing code that uses threads is different from most of what you have done in ECE 264. In particular, it can be challenging to debug. Give yourself time to understand multi-threaded programming.

BMPImage✶ median(const BMPImage✶ image, int radius, int num threads, const char✶✶ a error)

Q&A

1. Why can't we use division to calculate pixel intensities?
   This is to ensure that results will be predictable and consistent. If people used division, there might be subtle differences between implementations based on how you structure the calculation.
2. How can I calculate the color intensity without using division?
   First, let's look at how you would do this conceptually. You would convert each pixel to grayscale by computing the average of its red, blue, and green channels. That would yield a grayscale intensity between 0 and 255 for each pixel in the image. Then, for a given pixel, you would compare its grayscale intensity to the average of the grayscale intensities of the pixels in its neighborhood. Of course, computing these averages would use division. So how can we do it without division?
   Note that the sum of r+g+b for each cell is just 3 times the grayscale intensity. So, instead of dividing by 3, we will simply work with the r+g+b sum for every pixel. Likewise, instead of calculating the average within a neighborhood, we will multiply the value for the single pixel of interest by the number of pixels in its neighborhood and compare that with the sum of the r+g+b sums for every pixel in the neighborhood.
   Summary: Compare (r+g+b)*num_pixels_in_neighborhood with the sum of r+g+b over all cells in the neighborhood.
3. How can I compare two BMP files?
   
   For your unit tests, you can create a helper function, probably with a for loop that compares each pixel, one by one.
   For ad hoc checking, you can use command line tools. Suppose we have two files, expected.bmp and actual.bmp. The most basic thing to do would be to use xxd to store the hex dump of each and then diff them.

   xxd expected.bmp > expected.bmp.txt
   xxd   actual.bmp >   actual.bmp.txt
   diff expected.bmp.txt actual.bmp.txt
   

   We can do much better using a bash feature called process substitution, combined with a vim feature called vimdiff.

   vimdiff <(xxd expected.bmp) <(xxd actual.bmp)

   To skip the header, and compare just the pixel data, add -s +54 to each xxd command:

   vimdiff <(xxd -s +54 expected.bmp) <(xxd -s +54 actual.bmp)

   To look at only the header, add -l 54 to each xxd command:

   vimdiff <(xxd -l 54 expected.bmp) <(xxd -l 54 actual.bmp)

   To see the difference at the command line, instead of in vim, substitute diff in place of vimdiff in the above commands.
4. Why is the maximum number of threads 2,064,376?
   It is a system constant. To see the value on any Linux system, type this command in bash:

   cat /proc/sys/kernel/threads-max

Updates