Machine Vision - Practical 2

Last Updated : Thu 26/11/98 at 12:05pm

Important! This is the very latest version of this information page, and it includes information on image capture and preparation according to the most recent changes in the plans. If there is a conflict between the contents of this page and any other information source, please consider this page to be the most valid, until further notice.

Introduction

First of all, click here to see the definition of requirements for this practical. I need to remind you that this page is maintained by Professor Bernard Buxton and Professor Jan-Peter Muller, and is subject to amendments and modifications beyond my control.

Also, I will try to be available to answering questions most of the times beyond the scheduled practicals. Just e-mail me at: i.douros@cs.ucl.ac.uk

Now, some important admin information:

The deadline for this practical is Wednesday, 9/12/98. You should submit your reports to Professor Muller's office by 18:00 on that day. Please note that after that date Professor J.P. Muller will be away from the country, therefore an extension beyond that date cannot be given.

As from the 25th of November 1998 the practical is considered to be set. I suggest that you use your time wisely and take a look in the literarure recommended to you. Make sure you are familiar with the theory of the Canny Edge Detector, the Sobel mask functions and the Sonka multispectral edge detector before you start implementing these things. In addition, I will also provide you with some paper that should help you implement the Canny detector.

Image preparation

This section is to replace the instructions already provided to you about colour image conversion (which, if you want, you can access by clicking here.) These instructions require the use of HIPS library. To avoid using the HIPS library (which, as you already know, does not work anymore), follow these instructions instead.

First of all, use a camera and the capture application to grab your image. Save it as .rgb as usual.
Open your .rgb file with xv and save it as a full colour GIF. Let us assume, for convenience, that your file is called face.gif.
Now, you need to load the file in IDL. Remember, it is a colour image. To proceed, you also need to know the size of the file. Let us assume, for convenience, that our example file's size is 640x480 pixels.
Run IDL. At the prompt, type:

IDL > read_gif, 'face.gif', face, r, g, b

This will read the file, and will create three arrays. The array called face is 640x480, and each element is a number from 0 to 255. Note, however, that this number does not represent an intensity. It is the location of the colourmap where the colour coordinates are stored. The colourmap, in turn, is stored in arrays r, g, and b, one for each channel. Each one of these three arrays is 1D with 256 elements.
You can use these three arrays together with the face array in order to retrieve the three colour channels of the image. To do so is really easy. First, you have to declare the three arrays that will hold the colour channel information. Type:

IDL > channel_r = bindgen(640,480) IDL > channel_g = bindgen(640,480) IDL > channel_b = bindgen(640,480)
Finally, you need to pass the colour information to these three arrays. The most straightforward way to achieve this is the following:

IDL > for i=0,639 do for j=0,479 do channel_r(i,j)=r(face(i,j)) IDL > for i=0,639 do for j=0,479 do channel_g(i,j)=g(face(i,j)) IDL > for i=0,639 do for j=0,479 do channel_b(i,j)=b(face(i,j))
Your three channels are now ready for further processing.

Using the supplied implementation of the Canny Edge Detector

For assessing the quality of your results, an implementation of the Canny Edge Detector has been kindly supplied by Gideon Amos. To use it, you must first load the appropriate binary by typing the following at the IDL prompt:

IDL > restore, filename='~gideamos/pub/canny.dat'

Then, you can detect the edges of image by typing:

IDL > result = canny(image, gmask, lhthr, uhthr) IDL > tvscl, result, /order

Where:

gmask: width of the gaussian smoothing mask (divisible by 6)
lhthr: lower hysteresis threshold
uhthr: upper hysteresis threshold

In order to gain a better understanding of the Canny detector internals, it is advisable that you experiment with the values of these parameters and notice the differences in the final result.

Hints on the Canny Detector internals

To get started with the Canny implementation, you must keep in mind that the algorithm has three discrete stages. To avoid confusion, I suggest that you deal with them one at a time.

Note: I had promised to bring over a paper for you that contains the description of these steps. It appears that I don't have it anymore, so it might take another day to dig it out. Meanwhile, you can check out the information provided here.

So, the stages are:

Smoothing and differentiation

Convolve with a 2D Gaussian mask to eliminate high frequency noise
Calculate the partial derivatives of the image
Use them to calculate the gradient magnitude and gradient direction

Non-maximal suppresion

Find the pixels in the gradient magnitude that are not local maxima along the direction that is perpendicular to the edge.
Suppress them (set them to zero).

Hysteresis thresholding

Set two thresholds and threshold the suppresed result.
If the pixel is greater than the upper threshold, keep it.
If it's lower then the lower threshold, dump it.
If it's between the two thresholds, keep it only if it is 'connected' to a pixel that is greater than the upper threshold.

You should now have a fairly clear starting point in order to get on with this practical. Do not forget, that you can contact me at any time if you need help.

Good luck!

Yannis