Description
2 Background
The purpose of this assignment is to explore the use of shape and texture descriptors to classify images.
2.1 Assignment Synposis
In Question 1 you will classify leaf shapes described by the Histogram of Curvature Scale feature. In
Question 2 you will classify texture images using GLCM texture features and local binary patterns.
2.2 Datasets
Datasets for question 1
For question 1, you will use images similar to the ground truth images in assignment 3. You will download
two zip files:
leaftraining.zip These files naming scheme is threshimage_????.png where the wildcard characters
represent digits of a four-digit number. The first ten files (in lexicographic order) are of class 1, the
next ten files are of class 2, and the final ten files are of class 3. The images are binary PNG images.
leaftesting.zip This zip contains 129 test images. These are binary PNG images with filenames of the
form image_????.png where the wildcard characters represent digits of a four-digit number. The
first 50 files (in lexicographic order) are of class 1, the next 27 are of class 2, and remaining 52 are of
class 3.
Datasets for question 2
For Question 2 you will use images of texture. Each image contains the same texture throughout.
brodatztraining.zip This archive contains 15 training images for each of 8 texture classes (120 images
in total). These are PNG images with the names trainingPatch-??????.png where the wildcard
entries represent a six-digit number. The first digit of the number indicates the class label for the
training patch. The remaining digits are random. When you read these filenames with os.walk()
function, the first 15 files will be samples of texture class 1, the next 15 will be for texture class 2,
and so on. The training samples are unrotated samples of the texture classes.
brodatztesting.zip You are also provided with 40 test images for each of the 8 texture classes (320 images
in total). These are PNG images with the names patch-??????.png. Again, the wildcard entries are
a six-digit number. The first digit indicates the class label. Similar to the training data, when you
read the filenames with os.walk(), the first 40 elements of the array will be of texture class 1, the
next 40 will be of texture class 2, and so on. To make things interesting, each test image is a sample
of the original texture but rotated by a random angle between 0 and 360 degrees. This will enable
us to get a good feeling of how invariant to rotation our texture features are. Note: the test images
contain some imperfections that were deliberately left in the data set to make the problem a bit more
interesting and challenging (but not too challenging!).
2.3 Confusion Matrix
A confusion matrix is a quick way to visualize both how well a classifier is performing, and which pairs of
classes are causing the most problems (in terms of mis-classifications). It is considerably more informative
than the overall classification rate. The confusion matrix is a square matrix of size N where N is the
number of classes and entry (i, j) (row i, column j) is the number of times an image of class i was classified
as class j. Naturally then, if the classification rate is 100%, the confusion matrix is a diagonal matrix with
all off-diagonal entries equal to zero.
Example 1
In the following confusion matrix, most samples were classified correctly. 4 samples of class 1 were incorrectly
classified as class 2. 5 samples of class 6 were incorrectly classified as class 7, etc. At a glance, we can see that there
were a few instances of problems distinguishing between classes 1 and 2, 1 and 3, and 6 and 7. Additionally there
were some rare, isolated confusion between other classes.
1 2 3 4 5 6 7 8
1 35 4 0 0 0 0 0 0
2 0 36 0 0 0 0 0 0
3 4 0 39 0 0 0 0 1
4 0 0 0 39 0 0 0 0
5 0 0 0 0 40 0 0 0
6 0 0 1 0 0 40 5 0
7 1 0 0 1 0 0 35 0
8 0 0 0 0 0 0 0 39
2.4 Classification Rate
The classification rate is the percentage of images (feature vectors) that were correctly classified. In Example
1, there were a total of 17 misclassifications over 320 images, for a classification rate of 303/320 × 100% =
94.7%.
Page 2
3 Problems
Question 1 (29 points):
In this question you will classify pre-segmented images of leaves using the histogram of curvature
scale feature.
Detailed instructions are provided asn5-q1.ipynb. You should be able to get a classification rate of
more than 95%.
Step 2 Sample Output
If your output from step 2 looks like this, then your HoCS function is probably working.
0 5 10 15 20 25 30
0.0
0.2
0.4
0.6
0.8
1.0
Question 2 (29 points):
In this question you’ll classify images of various texture into eight different texture classes. Again we
will use a K-nearest neighbour classifier.
Detailed instructions are provided in asn5-q2.ipynb.
Using GLCM features, if you’re beating 65% correct classfication rate, you’re doing excellently. For
LBP/VAR8 features you should be able to get much higher; it shouldn’t be too hard to exceed 95%.
4 Files Provided
asn5-qX.ipynb: These are iPython notebooks, one for each question, which includes instructions and in
which you will do your assignment.
leaftraining: Folder containing training images for classifier in Question 1.
leaftesting: Folder containing testing images for classifier in Question 1.
brodatztraining: Folder containing training images for the classifiers in Question 2.
Page 3
brodatztesting: Folder containing testing images for the classifiers in Question 2.
5 What to Hand In
Hand in your completed iPython notebooks, one for each question.
6 Appendix A — Grading Rubric
The grading rubric can be found on Moodle.
Page 4
