nvidia-AI.summerschool17-no.../teaching-lw/ocr-hw/HOMEWORK
2017-09-22 19:26:46 +02:00

62 lines
2.4 KiB
Text

HOMEWORK
========
Most of the code for the following functions is already in the lecture notes. You
need to pull it out and put it together in some functions. You need no extra libraries
beyond what was covered in lecture.
1. Implement convolution using the fft2/ifft2 functions. Your function should look like
def fftconvolve(image,mask):
...
Image and mask are allowed to be different sizes, so you need to do something about the
mask. The easiest is to pad the mask so that it is the same size as the image.
2. Implement an image standardization function. That is, a function that, given
a rectangle size b=(bw,bh) replaces each pixel by subtracting the mean over that
rectangle from the pixel and then divides the result by the norm of the patch.
def standardize(image,b):
...
Implement the convolutions necessary for standardization using fftconvolve above.
3. Implement normalized cross correlation using your fftconvolve function and the
standardize function you defined above.
def ncc(image,template):
...
Output the image. Demonstrate your function on the w.png image and letter
templates you obtain using the charread.py file (you may want to look at it).
import charread
chars = charread.charread()
imshow(chars[17][0])
imshow(chars[17][5])
Note that you need to run ncc on page images with black background (why?).
OPTIONAL
=========
The following two problems are a lot more work, but quite rewarding.
4. Implement normalized cross correlation using multiple templates.
def nccs(image,templates):
...
This works like the function above but returns a list of normalized cross correlations,
one for each template. This makes the simplifying assumption that it is sufficient
to standardize the input image only once with parameters that are about right for all the
templates (use the mean width and mean height). Furthermore, the Fourier transform
of the standardized image for the convolution step needs to be carried out only once.
5. Compute normalized cross-correlations for all the templates and try to
label all the characters in the input image uniquely (i.e., so that there is only
one peak/output per input character). Remember that you know the approximate size
of all the characters, and they are all about the same size..
6. Try to find a way of outputting the characters in reading order (hard!).