GV12/3072 Coursework 1

This is the first coursework for course GV12/3072 on Image Processing. The subject of the coursework is simple image segmentation. For a basic pass in the coursework, you need to make a good attempt at the core section. To get full marks you need to have a good go at one or more of the additional sections.

I recommend you use matlab, but you are free to use any programming language you choose.

Hand in a short report on your work, which should be two written sides of A4 at most, plus any pictures and plots showing your results and a print out of your code. The short report should start with a brief list of what you have attempted and to what extent you have acheived each item. Next, write a short description of the methods you used for each part together with any conclusions you have drawn from your experiments.

The hand-in deadline for this coursework is 12:00 on 22nd October. Hand in your report to JJ or Tricia in the CS departmental office. Make sure you attach a completed coursework cover sheet to your work.

Core section

Here is a classic image processing test-image called girlface:

We would like to separate the skin region from the background of the image and surrounding parts of the head.

Download the image above and plot a histogram of the grey levels it contains.

Write a short program to threshold the image and try to identify a good threshold by trial and error.

Create a ground truth segmentation. If you are using Matlab, look up the function roipoly, which will allow you to do this easily.

Use your ground truth segmentation to plot an ROC curve for your thresholding algorithm. What threshold does the ROC curve suggest? How does it compare to your trial and error estimate?

NOTE: Feel free to use another image in addition to this one, but clearly define how that segmentation task has different challenges to the girlface.

Additional section

You do not need to do all of this, but to get more than a basic pass have a go at one part at least.

Examine the effects of noise on the performance of your segmentation algorithm. Add various levels of synthetic noise to the face image and see how the performance changes. You can choose from several types of noise, but be careful that the image grey levels remain valid after you add the synthetic noise.

Implement a region growing algorithm and compare its performance at segmenting the face in the image above with the performance of the thresholding algorithm.

Implement the ad-hoc algorithm for automatic threshold selection. How well does its suggestion match the thresholds chosen by trial and error and from the ROC curve?

Attempt to improve segmentation by first normalizing by a hand-made background. Where does this help or hurt?