I have a very large image and I want to scan it with a small template (T) and store in a database and use that database for image reconstruction. But, if I want to scan all of the image, the size of my pattern database will be very large. Is there any solution by which be able to reduce the size of pattern database? Or, is there any way to scan the image to make a smaller and yet rich pattern database?
Asked
Active
Viewed 369 times
1 Answers
2
You could run PCA on your template database to generate an alternate orthogonal template basis. All of your original templates are linear combinations of these basis templates. Since convolution is a linear operation you can convolve the image with each of the basis templates and recombine the results using the appropriate weights the get the results you would have gotten with each of the original templates.
PCA is usually used for dimensionality reduction, and so, you can choose just the first N basis templates as your approximate alternate templates. These should give a good approximation for the original DB but with reduced computational time.
N should be determined by the magnitude of the eigen-value corresponding to the selected basis eigen-vector (template).
Adi Shavit
- 16,743
- 5
- 67
- 137
-
@ Adi: thanks for your response. Is there any method better than PCA? I want to use another and better method – Sam Sep 04 '12 at 07:36
-
PCA is the method for building an orthogonal basis. – Adi Shavit Sep 04 '12 at 07:40