Hi all,
I have a quick question regarding template matching. I have one template image and another smaller input image which may be appeared in the template image with different size. For example, template size is NXN and the input image is BXB (B smaller than N) and the input image could be in template with any size different than BXB.
Cheers,
To match different size data templates you should index them to same size data pattern/templates.
To index them to same size data pattern/templates you can rely on spectral, temporal, spatial or salient features to extract the same size indexed data pattern/template.
You first have to answer the following questions (to yourself):
Can the smaller Image (needle) be scaled or rotated to fit in the larger one (haystack)?
Can the needle be modified in the RGB/ HSV dimensions to fit in the haystack?
Are you dealing with artificial Images containing letters or geometrical shapes or are you dealing with naturalistic photographs?
Thanks all,
There is no rotation between templates and there is no knowledge about the sizes. Images are natural images not artifical ones. It is kind of blind search and sliding windows so very tedious. That is why I am looking for a faster way.
U. Dreher, in the case of size-invariant features what do you suggest?
Could it be you are working in the face recognition/people tracking field?
As about size-invariant features: contour/'form' does not change with scaling. A triangle is a triangle irrespective of scale.
I can imagine other schemes for size-independent features, eg. the ratio between individual frequencies in the fourier coefficients of sub-images, but feel that the segmentation approach might serve your needs. Which is significantly easier to realize than an FT characterization approach.
Regards
If you re using natural images I suggest You start with comparing only a small set of points from your smaller image, maybe the four corners. If you find a corresponding set within the bigger image, you can check the rest. if not, you have saved some time.
You can use the correlation function to find the given template within an image.
corr2(A,B)==1,A=input image B=template
Template is found within an image
Keyvan
For finding the best match of a patch through an image, I found attached file few months ago and made some changes, it works well.
Good luck
You can use SIFT or SURF based matching. Once points are matched using SIFT or SURF, it is a common approach to use RANSAC in order to avoid mismatch or false matches. Hope that helps.
Basavaraj V Hiremath,
As far as I know, correlation only works for the cases in which sizes are similar.
the efficient algorithm to match different size sequences is DTW and time processing depends on the representation that you use.
Hi Keyvan.
U can walkthrough the window function(smaller size image) in an image(Bigger size).Wherever U find the correlation,there the particular image(template or window) exist.
SURF is one of the best for it. It is scale invariant method that can deal with input and template with different size
- Badges
- Science topic