What are invariances?
What is the effect of invariances on nearest neighbor classification?
What is the centroid of a pattern?
How do you compute the centroid of a pattern?
How can you use the centroid in order to improve classification?
How can you compensate for different translations of characters?
What is a canonical representative (of a class)?
What is an invariant representation?
What are Hu moments and why are they important?
How can you use the Fourier transform to achieve a translation invariant representatiobn?
How can you use the Fourier transform to achieve a rotation invariant representation?
What is the Fourier-Mellin transform?
What is a centralized second order moment?
What is skew correction and how is it computed?
State the equation for the coordinate transformation for skew correction.
* What is the mathematical definition of a norm/metric?
* What is the mathematical relationship between a norm and a metric?
* Write down the formulas for computing the Euclidean,
Manhattan, Taxicab, Hamming distance, maximum, Chebyshef, Cityblock,
and Minkowsky norm/metric.
* Give plausible examples of dissimilarity measures that violate each of the axioms.
* Consider a pattern recognition problem in which input vectors
v are to be classified as classes c. Explain how the problem is
affected if the input vector v is transformed into k v for some real
number k.
* Define a translation invariant metric.
* Explain what the effect of using a translation invariant metric is on a recognition problem.
* Explain how you can construct a useful translation invariant
metric from a metric that is not translation invariant by geometric
matching.
* Explain how you can construct a useful translation invariant
metric from a metric that is not translation invariant by centroid
matching.
* State the formula for the centroid.
* Write loop-free NumPy code for computing the centroid of an image.
* Images of handwritten characters are sometimes encoded with
black=0/white=255, sometimes with black=255/white=0, and sometimes with
black=1/white=0. Explain what the effect of each choice is on the
centroid computation.
* What is the canonical representative of a pattern, given a set of possible transformations?
* What is an equivalence class of patterns under a transformation group (e.g. translation)?
* Explain anisotropic size normalization of characters by their bounding box.
* What is the difference between anisotropic and isotropic rescaling?
* What is the difference between finding a canonical
representative under translation using centroids, medioids, and
bounding boxes? Which one works best in practice?