GV12/3072 Coursework 3

This is the third coursework for course GV12/3072 on Image Processing. This coursework is on edge detection. For a basic pass in the coursework (50%), you need to do the core section. To get full marks (100%) you must succeed in completing two of the additional sections.

I recommend that you use matlab.

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. Make sure it is clear what commands you used to generate each of the pictures you include. Indicate the kernel values and any other parameters used for each figure.

The definitive hand-in deadline for this coursework is 12:00 (noon) on 1st December. 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 the test image:

We would like to locate the edges in the image as consistently and accurately as possible. The core part is:

Compute edge maps (ie follow the "Simple edge-detector" procedure from the lecture notes) for the image using YOUR OWN (a) Sobel edge detector (Sobel filters and thresholding), (b) Prewitt edge detector (Prewitt filters and thresholding), (c) an edge detector using derivative of Gaussian filtering in place of the Sobel/Prewitt kernels and (d) the built-in Canny edge detector implemented in matlab. Compare the output visually and comment on which you think is best and why. List the tunable parameters in each edge detector and explain how you chose settings for each. Note: the default parameter settings are not necessarily the best, and your results may differ from the built-in Canny function.

Implement a non-maximal suppression algorithm and use it on the responses of the filters you used in (c) above.

Implement hysteresis thresholding and generate binary edge maps from the non-maximal suppression results obtained above. One possibility is to use bwselect(). Here, you can ignore orientation. Describe how you choose settings for the thresholds in your algorithm. Comment on how the results compare with matlab's Canny edge detector.

Additional section

You do not need to do all of these, but must correctly complete two of them to reach full credit.

Implement an alternative hysteresis algorithm that uses the edge orientation in each pixel in the test to determine whether it has a neighbour already in the edge map. Does the adaptation improve results?

The use of a scale space pyramid can improve on standard hysteresis. Rather than starting with edge maps from high and low thresholds, start with edge maps from high and low standard deviation in your Gaussian kernels. What differences do you observe? Extend your hysteresis to use the whole series of edge maps from increasing standard deviations in the Gaussian kernels.

Construct a ground truth edge map by labelling just the main edges of the butterfly's wings (or more detail, if you feel like it) in the following image by hand. Construct ROC curves for your version of the Canny edge detector and compare its performance to the built-in Canny. Explain how you classify pixels in your edge maps as true/false positives/negatives.

The alternative to the ROC that may be more suitable for evaluating edge detectors is the precision-recall (PR) curve. Precision (vertical axis) is the fraction of detected edges that have a matching edge in the ground truth. Recall (horizontal axis) is the fraction of ground-truth edges that are detected. Plot the PR curve using a simple matching procedure that simply marks edge pixels in one map as matched if there is an edge pixel within distance D in the other image; varying D produces the PR curve. Hint: use the distance transform.