Minutiae points detection and matching

I have been working on minutiae points detection using template matching. The goal was to find a filter that, when applied on the enhanced print, pops out the bifurcation and ridge endings.

After trying some real minutia point template in various direction, I decided to create a template using maths functions. The main motivation is because it allows to orient the template in any direction without problems in filling the holes of the rotation. It is also convenient because it can be tuned easily. I did not remember having seen this kind of approach in the literature or the net, but it is attractive and quite powerful. I used interpolated cosine wave functions with different frequencies.

Due to the duality of ridge ending and bifurcation, the enhanced image has to be filtered using two filters. This can be done by inverting the image, or the filter, or rotating the filter of 180°. I chose the first solution even if it is not the faster one. Then the local maxima of the two filtered images need to be extracted and merged together (a threshold can be adjusted here). The next picture shows all those steps.



After having the minutiae point position and direction, I tryed Ransac on different prints even if I do not specify correspondences. The next picture shows that the two fingerprints are considered to match.



I also made functions to save and load minutiae points from a text file, which allows me to begin some tests. The first impression is that is is very slow, because I compute for each pixel the corresponding oriented Gabor filter and minutia template. It takes 15 seconds to enroll a print. It also take 10 seconds to do 5000 iterations of Ransac. An other thing that can be improved is to use pointers and increment them instead of using the convenient function cvGetReal2D(img, i, j). The second impression is that I do not have false accepts, but some true rejects. That is because I mainly tune the program on good quality prints. Also Ransac need correspondences to work accurately and decrease those rates.

Next steps:
- Find correspondences: I tryed to run David Lowe's SIFT algorithm, but it gives features of 1 pixel size. Moreover, It founds more than 2000 features. I may also try Normalized Cross Correlation.
- Do the draft report.
- Test and tune the program.